{"slug": "gpt-5", "name": "GPT-5", "category": "ai", "category_name": "AI models and assistants", "description": "OpenAI's GPT-5 family of large language models, available through ChatGPT, Codex and the API.", "recommenders": 113, "commenters": 205, "warners": 33, "score": 82.03, "rising": -7.44, "first": "2025-11-04", "latest": "2026-10-20", "why": "Liked for capable coding work and close instruction following, with several commenters switching from Claude and citing roomier usage limits and good token efficiency.", "why_is_quote": false, "has_verdict": true, "sentiment": {"recommend": 112, "neutral": 49, "compare": 27, "warn": 36}, "spark": [4, 4, 6, 2, 4, 8, 6, 29, 29, 10, 8, 2], "listed": true, "loved": ["Coding ability", "Cost and limits"], "caveat": "Stops early", "short_quote": "GPT-5.6 so far has been a significantly better technical writer in my opinion, and it's much faster in conversations.", "icon": true, "domain": "senko.net", "github_repo": null, "official_url": "https://openai.com", "pricing": "freemium", "aliases": ["GPT-5.5", "GPT 5.5", "GPT-5.6", "GPT-5.4", "GPT-5", "GPT-5.2", "GPT-5.1", "GPT 5.6", "GPT 5.4", "GPT 5.2", "GPT 5.1", "GPT 5.x", "GPT5.6", "GPT-5.x"], "verdict": [{"text": "Liked for capable coding work and close instruction following", "cites": [{"id": 781598688, "item_id": 48849918, "url": "https://news.ycombinator.com/item?id=48849918", "by": "edumucelli", "date": "2026-07-09", "stance": "recommend", "quote": "GPT won hands down.", "excerpt": "Not sure about the consensus, but during an entire week I have done every task on my workplace with both Opus 4.8 and GPT 5.5. GPT won hands down. I would even sometimes copy the plans and solutions (using different Git worktrees) from GPT and paste it on Opus and itself would say GPT plans were better. At that point I have migrated. Fable is not enabled in our workspace so I have not tried. Claude lost my trust arou", "context": "", "use_case": "workplace coding tasks", "switched_from": "Claude", "story": "GPT-5.6"}, {"id": 788852401, "item_id": 49303275, "url": "https://news.ycombinator.com/item?id=49303275", "by": "jmstfv", "date": "2026-08-14", "stance": "recommend", "quote": "Both GPT 5.5 and 5.6 are quite capable", "excerpt": "I've switched to Codex a few months ago when Claude's weekly limits were getting pretty stiff, and I haven't looked back. Both GPT 5.5 and 5.6 are quite capable, especially compared to nerfed Opus 4.7 (haven't tried 5). Also, the Codex guy regularly resets weekly limits for everyone, which is a nice bonus (I know it's a temporary gimmick to attract more users, but I might as well use it while it lasts.)", "context": "", "use_case": "AI coding", "switched_from": "", "story": "Why does Opus 5 feel worse to work with?"}, {"id": 795703408, "item_id": 49731463, "url": "https://news.ycombinator.com/item?id=49731463", "by": "MAZ97", "date": "2026-09-16", "stance": "recommend", "quote": "Gpt 5.6, it understands me perfectly and is able to always do what i tell it to the point", "excerpt": "Gpt 5.6, it understands me perfectly and is able to always do what i tell it to the point", "context": "", "use_case": "", "switched_from": "", "story": "Ask HN: What default model do you use and why?"}, {"id": 784010610, "item_id": 49000663, "url": "https://news.ycombinator.com/item?id=49000663", "by": "caspar", "date": "2026-07-21", "stance": "recommend", "quote": "GPT 5.4/5.5 would prefer them.", "excerpt": "This is partially a \"how much does the model follow instructions\" thing. I use Pi with a vibed LSP extension and Claude (4.6 or thereabouts) almost never followed instructions to use LSP renaming tools - despite it being strongly emphasized in system prompt and agents.md. However I found Codex 5.3 would use them sometimes, and GPT 5.4/5.5 would prefer them.", "context": "", "use_case": "LSP refactoring", "switched_from": "", "story": "Stop Using OpenCode"}]}, {"text": ", with several commenters switching from Claude and citing roomier usage limits and good token efficiency.", "cites": [{"id": 789504304, "item_id": 49344019, "url": "https://news.ycombinator.com/item?id=49344019", "by": "justinbaker84", "date": "2026-08-18", "stance": "recommend", "quote": "I switched from claude to gpt when 5.6 came out for the same reasons.", "excerpt": "I switched from claude to gpt when 5.6 came out for the same reasons. I don't understand why so many people are still using claude when GPT and open source are so much better.", "context": "", "use_case": "AI assistant / coding", "switched_from": "Claude", "story": "GPT-5.6 Sol Pricing Cut by 50%"}, {"id": 781591552, "item_id": 48849472, "url": "https://news.ycombinator.com/item?id=48849472", "by": "jakswa", "date": "2026-07-09", "stance": "recommend", "quote": "With gpt 5.5 I don't think I'm working differently but haven't felt the need to check anything because I think I've hit my limit... once?", "excerpt": "I'll agree and expand on \"weird restrictions\" -- I used to check the claude usage graphs multiple times a day to see where I'm at on my weekly budget. With gpt 5.5 I don't think I'm working differently but haven't felt the need to check anything because I think I've hit my limit... once? on some egregious edge case scenario iirc", "context": "", "use_case": "coding", "switched_from": "Claude", "story": "GPT-5.6"}, {"id": 792321072, "item_id": 49520067, "url": "https://news.ycombinator.com/item?id=49520067", "by": "ramon156", "date": "2026-09-01", "stance": "recommend", "quote": "the usage and quality out of gpt 5.6 is on par, if not better in token usage.", "excerpt": "the usage and quality out of gpt 5.6 is on par, if not better in token usage. I would love to do a write-up on which tools are useful for what use-case, as AA keeps improving. I use gemini for rewriting code docs, because the frontier models are so verbose when it comes to writing text", "context": "", "use_case": "coding", "switched_from": "", "story": "Is it just me, or has Claude Opus gotten worse recently?"}, {"id": 781862193, "item_id": 48866387, "url": "https://news.ycombinator.com/item?id=48866387", "by": "arendtio", "date": "2026-07-10", "stance": "recommend", "quote": "the GPT 5.6 family does a good job in terms of cost for performance", "excerpt": "I also like to take a look at https://cursor.com/cursorbench While in the past months Composer 2.5 was a lot better than I had expected a year ago, and the GPT 5.6 family does a good job in terms of cost for performance, I wonder why nobody is talking about Grok 4.5 high? Those numbers look very convincing to me.", "context": "", "use_case": "coding", "switched_from": "", "story": "GPT-5.6"}]}, {"text": " The most common complaint is that 5.6 releases burn more tokens than 5.5", "cites": [{"id": 785781250, "item_id": 49111328, "url": "https://news.ycombinator.com/item?id=49111328", "by": "runjake", "date": "2026-07-30", "stance": "warn", "quote": "5.6 seems to use too many tokens to do the same or slightly inferior job", "excerpt": "Sonnet 5 Medium and GPT 5.5 Medium for coding. I drop down to a Mini model for file-based tasks. Codex keeps trying to revert me back to 5.6 but 5.6 seems to use too many tokens to do the same or slightly inferior job. I am only on multiple $20 plans.", "context": "on multiple $20 plans", "use_case": "coding", "switched_from": "", "story": "Ask HN: Which one do you use for planning and coding between sonnet and Opus?"}, {"id": 787860976, "item_id": 49241311, "url": "https://news.ycombinator.com/item?id=49241311", "by": "vincent_s", "date": "2026-08-10", "stance": "warn", "quote": "It definitely feels like subscription limits have dropped since GPT-5.6 came out.", "excerpt": "It definitely feels like subscription limits have dropped since GPT-5.6 came out. But I couldn't really find any evidence for it, going through my history. All I could come up with is that 5.6-sol uses about twice the tokens compared to 5.5 (both on xhigh) [0]. One thing I just did was to stop four long-running Codex sessions that ran on 5.6-sol and switched to 5.5 and asked it \"Please check if you're really going ag", "context": "", "use_case": "coding", "switched_from": "", "story": "Switching from GPT-5.5 to GPT-5.6 Made Me Less Productive"}, {"id": 792877042, "item_id": 49554815, "url": "https://news.ycombinator.com/item?id=49554815", "by": "wahnfrieden", "date": "2026-09-03", "stance": "recommend", "quote": "whereas I was able to work ~80 hours/week with 5.5 High", "excerpt": "2.5x more expensive than Sol. Can expect 2.5x more usage in Codex subscription. Sol is already brutal (even after their recent fixes, it's just a token-hungry model: I go through a full 20x account per day, on Sol Med/High standard speed, with ~2 threads). Note that Tibo recommended using Sol Med as daily driver. When I'm doing less complicated work, I can't even make it past 2-3 days with Sol Med, whereas I was able", "context": "", "use_case": "coding", "switched_from": "", "story": "OpenAI begins rolling out GPT-6 Astra"}]}, {"text": ", and that the models stop early, give up or drift on long tasks.", "cites": [{"id": 786465761, "item_id": 49154110, "url": "https://news.ycombinator.com/item?id=49154110", "by": "sdoering", "date": "2026-08-03", "stance": "recommend", "quote": "switching back to gpt-5.5 for \"real work\"", "excerpt": "In my daily work, I found expeciall terra and sol now stopping every few rounds again, telling me the tak is done. I even had them create a detailled plan - and told them to finish \"end to end\" - and they appruptly stop after the plan. Because they interpret this as finished. Even if the DOD is clearly not \"finish the plan\". The new models are shite (pardon my French), when it comes to long running tasks and I find m", "context": "", "use_case": "real work", "switched_from": "GPT-5.6 Sol", "story": "The Shape of Things to Come"}, {"id": 794782417, "item_id": 49673901, "url": "https://news.ycombinator.com/item?id=49673901", "by": "gpugreg", "date": "2026-09-12", "stance": "warn", "quote": "they often give up saying it can't be done, or cheat by modifying the tests", "excerpt": "I currently use DeepSeek-V4.1-Flash and the V4-Pro and -Flash versions before that, because they are extremely cheap. I have spent less than $25 for over a billion token so far (1B cached, 15M out, 19M in). I even prefer the DeepSeek models over the older OpenAI offerings. When I tell a GPT-5.x model to do some difficult task, they often give up saying it can't be done, or cheat by modifying the tests, while the Deep", "context": "", "use_case": "difficult tasks", "switched_from": "", "story": "Ask HN: What default model do you use and why?"}, {"id": 793242850, "item_id": 49577678, "url": "https://news.ycombinator.com/item?id=49577678", "by": "Topfi", "date": "2026-09-05", "stance": "warn", "quote": "GPT-5.5 simply did not stay on track post compaction", "excerpt": "Far too early for any real opinion, but GPT-6 Astra (on Light) stops a lot and often in unintuitive ways that I haven't seen in a while. Had a very hard to reproduce windows focus bug on macOS that was unreliable to reproduce and when it could be reproduced, it disappeared the second I switched to the app formerly known as Codex, so for the first time used the voice interface with that little hovering bubble, telling", "context": "", "use_case": "", "switched_from": "", "story": "Ask HN: Initial Thoughts on GPT-6 Astra?"}, {"id": 787860976, "item_id": 49241311, "url": "https://news.ycombinator.com/item?id=49241311", "by": "vincent_s", "date": "2026-08-10", "stance": "warn", "quote": "It definitely feels like subscription limits have dropped since GPT-5.6 came out.", "excerpt": "It definitely feels like subscription limits have dropped since GPT-5.6 came out. But I couldn't really find any evidence for it, going through my history. All I could come up with is that 5.6-sol uses about twice the tokens compared to 5.5 (both on xhigh) [0]. One thing I just did was to stop four long-running Codex sessions that ran on 5.6-sol and switched to 5.5 and asked it \"Please check if you're really going ag", "context": "", "use_case": "coding", "switched_from": "", "story": "Switching from GPT-5.5 to GPT-5.6 Made Me Less Productive"}]}], "praise": [{"label": "Coding ability", "summary": "Commenters find it writes clean, reliable code and compares well with Claude and Composer.", "quotes": [{"id": 797664129, "item_id": 49854008, "url": "https://news.ycombinator.com/item?id=49854008", "by": "epihelix", "date": "2026-09-26", "stance": "recommend", "quote": "It writes very nice, clean code with almost no comments.", "excerpt": "Yes, with Enterprise M365 Copilot. Select \"GPT 5.6 Think deeper\" (ugh, so unrevealing, MS) and you've got free, unlimited Sol or Terra (likely a mix depending on load) on tap. It writes very nice, clean code with almost no comments. I use it a lot.", "context": "", "use_case": "writing code", "switched_from": "", "story": "Microsoft abandons personal AI chatbot race with Copilot reboot"}, {"id": 781598688, "item_id": 48849918, "url": "https://news.ycombinator.com/item?id=48849918", "by": "edumucelli", "date": "2026-07-09", "stance": "recommend", "quote": "GPT won hands down.", "excerpt": "Not sure about the consensus, but during an entire week I have done every task on my workplace with both Opus 4.8 and GPT 5.5. GPT won hands down. I would even sometimes copy the plans and solutions (using different Git worktrees) from GPT and paste it on Opus and itself would say GPT plans were better. At that point I have migrated. Fable is not enabled in our workspace so I have not tried. Claude lost my trust arou", "context": "", "use_case": "workplace coding tasks", "switched_from": "Claude", "story": "GPT-5.6"}, {"id": 781370465, "item_id": 48835654, "url": "https://news.ycombinator.com/item?id=48835654", "by": "bhouston", "date": "2026-07-08", "stance": "recommend", "quote": "It is unusable compared to GPT 5.4 or 5.5.", "excerpt": "Every time I use Composer 2.5 I have to spend a bunch of time cleaning up its mistakes. It is unusable compared to GPT 5.4 or 5.5. My time is more valuable that I will use a model that doesn\u2019t f** up my code base.", "context": "", "use_case": "AI coding", "switched_from": "", "story": "Grok 4.5"}]}, {"label": "Cost and limits", "summary": "Users report good cost for performance, token efficiency and rarely hitting subscription limits.", "quotes": [{"id": 781862193, "item_id": 48866387, "url": "https://news.ycombinator.com/item?id=48866387", "by": "arendtio", "date": "2026-07-10", "stance": "recommend", "quote": "the GPT 5.6 family does a good job in terms of cost for performance", "excerpt": "I also like to take a look at https://cursor.com/cursorbench While in the past months Composer 2.5 was a lot better than I had expected a year ago, and the GPT 5.6 family does a good job in terms of cost for performance, I wonder why nobody is talking about Grok 4.5 high? Those numbers look very convincing to me.", "context": "", "use_case": "coding", "switched_from": "", "story": "GPT-5.6"}, {"id": 792321072, "item_id": 49520067, "url": "https://news.ycombinator.com/item?id=49520067", "by": "ramon156", "date": "2026-09-01", "stance": "recommend", "quote": "the usage and quality out of gpt 5.6 is on par, if not better in token usage.", "excerpt": "the usage and quality out of gpt 5.6 is on par, if not better in token usage. I would love to do a write-up on which tools are useful for what use-case, as AA keeps improving. I use gemini for rewriting code docs, because the frontier models are so verbose when it comes to writing text", "context": "", "use_case": "coding", "switched_from": "", "story": "Is it just me, or has Claude Opus gotten worse recently?"}, {"id": 781591552, "item_id": 48849472, "url": "https://news.ycombinator.com/item?id=48849472", "by": "jakswa", "date": "2026-07-09", "stance": "recommend", "quote": "With gpt 5.5 I don't think I'm working differently but haven't felt the need to check anything because I think I've hit my limit... once?", "excerpt": "I'll agree and expand on \"weird restrictions\" -- I used to check the claude usage graphs multiple times a day to see where I'm at on my weekly budget. With gpt 5.5 I don't think I'm working differently but haven't felt the need to check anything because I think I've hit my limit... once? on some egregious edge case scenario iirc", "context": "", "use_case": "coding", "switched_from": "Claude", "story": "GPT-5.6"}]}, {"label": "Instruction following", "summary": "Praised for doing what it is told and adhering to tasks and tool instructions.", "quotes": [{"id": 795703408, "item_id": 49731463, "url": "https://news.ycombinator.com/item?id=49731463", "by": "MAZ97", "date": "2026-09-16", "stance": "recommend", "quote": "Gpt 5.6, it understands me perfectly and is able to always do what i tell it to the point", "excerpt": "Gpt 5.6, it understands me perfectly and is able to always do what i tell it to the point", "context": "", "use_case": "", "switched_from": "", "story": "Ask HN: What default model do you use and why?"}, {"id": 793242849, "item_id": 49577678, "url": "https://news.ycombinator.com/item?id=49577678", "by": "Topfi", "date": "2026-09-05", "stance": "recommend", "quote": "GPT-5.4 (which remains the best model by any lab I ever tested for task adherence", "excerpt": "Far too early for any real opinion, but GPT-6 Astra (on Light) stops a lot and often in unintuitive ways that I haven't seen in a while. Had a very hard to reproduce windows focus bug on macOS that was unreliable to reproduce and when it could be reproduced, it disappeared the second I switched to the app formerly known as Codex, so for the first time used the voice interface with that little hovering bubble, telling", "context": "", "use_case": "task adherence", "switched_from": "", "story": "Ask HN: Initial Thoughts on GPT-6 Astra?"}, {"id": 784010610, "item_id": 49000663, "url": "https://news.ycombinator.com/item?id=49000663", "by": "caspar", "date": "2026-07-21", "stance": "recommend", "quote": "GPT 5.4/5.5 would prefer them.", "excerpt": "This is partially a \"how much does the model follow instructions\" thing. I use Pi with a vibed LSP extension and Claude (4.6 or thereabouts) almost never followed instructions to use LSP renaming tools - despite it being strongly emphasized in system prompt and agents.md. However I found Codex 5.3 would use them sometimes, and GPT 5.4/5.5 would prefer them.", "context": "", "use_case": "LSP refactoring", "switched_from": "", "story": "Stop Using OpenCode"}]}, {"label": "Writing style", "summary": "Commenters like its direct communication and technical writing, especially with plain-English prompts.", "quotes": [{"id": 788015744, "item_id": 49250984, "url": "https://news.ycombinator.com/item?id=49250984", "by": "ferris-booler", "date": "2026-08-10", "stance": "recommend", "quote": "GPT-5.6 so far has been a significantly better technical writer in my opinion, and it's much faster in conversations.", "excerpt": "Yeah I hit this feeling with Claude one too many times and switched from Anthropic Pro to OpenAI Pro. GPT-5.6 so far has been a significantly better technical writer in my opinion, and it's much faster in conversations. My guess is that Claude's fantasy-jargon is a symptom of training failure, not a sneaky intelligence edge (I could be wrong). Tangential but I use OpenCode with GPT-5.6 rather than Codex because I cou", "context": "", "use_case": "technical writing and conversation", "switched_from": "Claude", "story": "Humanising LLM Outputs Is Dumb"}, {"id": 789512418, "item_id": 49344526, "url": "https://news.ycombinator.com/item?id=49344526", "by": "Aperocky", "date": "2026-08-18", "stance": "recommend", "quote": "The communication is more direct too with GPT5.6 family of LLM.", "excerpt": "One of the biggest reason I migrated to Codex. I'd like to see each agentic CLI process as a simple building block, and having 100 claude session means I sometimes run out of memory (and I have 256GB of them). Codex never have this problem. The communication is more direct too with GPT5.6 family of LLM. No more belt and suspenders.", "context": "", "use_case": "coding", "switched_from": "", "story": "Linux 7.3 improves performance when running out of vRAM"}, {"id": 791494016, "item_id": 49468376, "url": "https://news.ycombinator.com/item?id=49468376", "by": "FooBarWidget", "date": "2026-08-27", "stance": "recommend", "quote": "With GPT-5.6 I get reasonable results with \"use natural, plain English\".", "excerpt": "I think this is more Claude writing than LLM writing. GPT-5.6 also talks like an LLM, but not like this. With GPT-5.6 I get reasonable results with \"use natural, plain English\". I tried \"use ASD-STE100 Simplified Technical English\" but I've found \"natural plain English\" to work better. Still not on the level I would write, but better. In the mean time, I hear that Claude resists changing its writing style. A PO used ", "context": "", "use_case": "writing in plain English", "switched_from": "", "story": "Show HN: Restoredrill \u2013 proves your Postgres backups restore"}]}, {"label": "Long context", "summary": "Some users notice no quality loss near context limits or after compaction.", "quotes": [{"id": 783508945, "item_id": 48969309, "url": "https://news.ycombinator.com/item?id=48969309", "by": "scotty79", "date": "2026-07-19", "stance": "recommend", "quote": "I use GLM5.2 and gpt-5.5 and 5.6 and never noticed any quality degradation near the limits of the context or due to compaction.", "excerpt": "OpenAI compaction is very different than Anthropic compaction, especially in 5.6 I use GLM5.2 and gpt-5.5 and 5.6 and never noticed any quality degradation near the limits of the context or due to compaction.", "context": "", "use_case": "AI coding", "switched_from": "", "story": "OpenAI reduces Codex Model Context Size from 372k to 272k"}, {"id": 784061296, "item_id": 49003831, "url": "https://news.ycombinator.com/item?id=49003831", "by": "edg5000", "date": "2026-07-22", "stance": "recommend", "quote": "I always let GPT 5.5/5.6 sol reach compatction.", "excerpt": "I always let GPT 5.5/5.6 sol reach compatction. I'm only offered a 258K window, but I think once they release 1M to normal plan users, it will be usable across the whole 1M. At least with Opus I could use the whole 1M. I'd argue that the more context I use, the better performance gets. It has more info already available. I don't notice degradation.", "context": "", "use_case": "long-context coding", "switched_from": "", "story": "Agent swarms and the new model economics"}]}, {"label": "Thoroughness", "summary": "Trusted for thorough work such as reverse engineering and porting legacy software.", "quotes": [{"id": 786628016, "item_id": 49164251, "url": "https://news.ycombinator.com/item?id=49164251", "by": "Tiberium", "date": "2026-08-03", "stance": "recommend", "quote": "I trust GPT models the most for reverse engineering, they're very thorough.", "excerpt": "Honestly such decomp projects are the place where LLM agents help the most, I have a similar one for a 1999 game which started from the agent renaming all functions, globals in the IDA DB, adding types and function signatures. After the database was basically 100% annotated, it was exported, made compilable with the same compiler as the original game. Then after some fixes (basically all being IDA decompilation bugs)", "context": "", "use_case": "reverse engineering decompilation", "switched_from": "", "story": "KisakCOD \u2013 open-source reimplementation of Call of Duty 4 Multiplayer"}, {"id": 782517217, "item_id": 48907326, "url": "https://news.ycombinator.com/item?id=48907326", "by": "tda", "date": "2026-07-14", "stance": "recommend", "quote": "Fable and GPT-5.6 chewed through the compiler warnings, dependency hell, and portability bugs largely on their own.", "excerpt": "I ported a decades-old piece of scientific software (Delft3D \u2014 Deltares' Fortran hydrodynamics engine) to run natively on Apple Silicon. From my phone, mostly, because I could (and because AI could). Runs pretty decently on a Mac Studio M3 now. It was officially Intel-compiler + Windows/Linux only; the port gets it building with gfortran on macOS/arm64, PETSc+MPI solver and all. Fable and GPT-5.6 chewed through the c", "context": "not a C or Fortran dev", "use_case": "porting software to Apple Silicon", "switched_from": "", "story": "Ask HN: What Are You Working On? (July 2026)"}]}], "caveats": [{"label": "Stops early", "summary": "Models sometimes give up, declare tasks done prematurely or drift on long-running work.", "quotes": [{"id": 786465761, "item_id": 49154110, "url": "https://news.ycombinator.com/item?id=49154110", "by": "sdoering", "date": "2026-08-03", "stance": "recommend", "quote": "switching back to gpt-5.5 for \"real work\"", "excerpt": "In my daily work, I found expeciall terra and sol now stopping every few rounds again, telling me the tak is done. I even had them create a detailled plan - and told them to finish \"end to end\" - and they appruptly stop after the plan. Because they interpret this as finished. Even if the DOD is clearly not \"finish the plan\". The new models are shite (pardon my French), when it comes to long running tasks and I find m", "context": "", "use_case": "real work", "switched_from": "GPT-5.6 Sol", "story": "The Shape of Things to Come"}, {"id": 794782417, "item_id": 49673901, "url": "https://news.ycombinator.com/item?id=49673901", "by": "gpugreg", "date": "2026-09-12", "stance": "warn", "quote": "they often give up saying it can't be done, or cheat by modifying the tests", "excerpt": "I currently use DeepSeek-V4.1-Flash and the V4-Pro and -Flash versions before that, because they are extremely cheap. I have spent less than $25 for over a billion token so far (1B cached, 15M out, 19M in). I even prefer the DeepSeek models over the older OpenAI offerings. When I tell a GPT-5.x model to do some difficult task, they often give up saying it can't be done, or cheat by modifying the tests, while the Deep", "context": "", "use_case": "difficult tasks", "switched_from": "", "story": "Ask HN: What default model do you use and why?"}, {"id": 793242850, "item_id": 49577678, "url": "https://news.ycombinator.com/item?id=49577678", "by": "Topfi", "date": "2026-09-05", "stance": "warn", "quote": "GPT-5.5 simply did not stay on track post compaction", "excerpt": "Far too early for any real opinion, but GPT-6 Astra (on Light) stops a lot and often in unintuitive ways that I haven't seen in a while. Had a very hard to reproduce windows focus bug on macOS that was unreliable to reproduce and when it could be reproduced, it disappeared the second I switched to the app formerly known as Codex, so for the first time used the voice interface with that little hovering bubble, telling", "context": "", "use_case": "", "switched_from": "", "story": "Ask HN: Initial Thoughts on GPT-6 Astra?"}]}, {"label": "Token hunger", "summary": "GPT-5.6 is reported to use more tokens than 5.5 for similar or slightly worse results.", "quotes": [{"id": 785781250, "item_id": 49111328, "url": "https://news.ycombinator.com/item?id=49111328", "by": "runjake", "date": "2026-07-30", "stance": "warn", "quote": "5.6 seems to use too many tokens to do the same or slightly inferior job", "excerpt": "Sonnet 5 Medium and GPT 5.5 Medium for coding. I drop down to a Mini model for file-based tasks. Codex keeps trying to revert me back to 5.6 but 5.6 seems to use too many tokens to do the same or slightly inferior job. I am only on multiple $20 plans.", "context": "on multiple $20 plans", "use_case": "coding", "switched_from": "", "story": "Ask HN: Which one do you use for planning and coding between sonnet and Opus?"}, {"id": 787860976, "item_id": 49241311, "url": "https://news.ycombinator.com/item?id=49241311", "by": "vincent_s", "date": "2026-08-10", "stance": "warn", "quote": "It definitely feels like subscription limits have dropped since GPT-5.6 came out.", "excerpt": "It definitely feels like subscription limits have dropped since GPT-5.6 came out. But I couldn't really find any evidence for it, going through my history. All I could come up with is that 5.6-sol uses about twice the tokens compared to 5.5 (both on xhigh) [0]. One thing I just did was to stop four long-running Codex sessions that ran on 5.6-sol and switched to 5.5 and asked it \"Please check if you're really going ag", "context": "", "use_case": "coding", "switched_from": "", "story": "Switching from GPT-5.5 to GPT-5.6 Made Me Less Productive"}, {"id": 792877042, "item_id": 49554815, "url": "https://news.ycombinator.com/item?id=49554815", "by": "wahnfrieden", "date": "2026-09-03", "stance": "recommend", "quote": "whereas I was able to work ~80 hours/week with 5.5 High", "excerpt": "2.5x more expensive than Sol. Can expect 2.5x more usage in Codex subscription. Sol is already brutal (even after their recent fixes, it's just a token-hungry model: I go through a full 20x account per day, on Sol Med/High standard speed, with ~2 threads). Note that Tibo recommended using Sol Med as daily driver. When I'm doing less complicated work, I can't even make it past 2-3 days with Sol Med, whereas I was able", "context": "", "use_case": "coding", "switched_from": "", "story": "OpenAI begins rolling out GPT-6 Astra"}]}, {"label": "Refusals", "summary": "Some users hit refusals on security scanning and certain tool-building tasks.", "quotes": [{"id": 784869634, "item_id": 49054352, "url": "https://news.ycombinator.com/item?id=49054352", "by": "alexsmirnov", "date": "2026-07-25", "stance": "warn", "quote": "gpt-5.6, gemini pro, and opus all rejected to follow.", "excerpt": "I did run both kimi-k3 and glm-5.2 with Capital One vulnhunt [1]. No rejections, they did find the same problems on my project that I used for testing. gpt-5.6, gemini pro, and opus all rejected to follow. gpt even declined to edit skill files. [1] https://www.capitalone.com/tech/open-source/announcing-vulnhunter/", "context": "", "use_case": "vulnerability hunting", "switched_from": "", "story": "UK AISI / Caisi Preliminary Assessment of Kimi K3's Cyber Capabilities"}, {"id": 781743682, "item_id": 48858980, "url": "https://news.ycombinator.com/item?id=48858980", "by": "sdoering", "date": "2026-07-10", "stance": "warn", "quote": "like gpt-5.5 not willing to build a downloading tool for Annas Archive", "excerpt": "To me, the question of switching (and any recommendation) depends highly on the type of work you do as well as your setup in terms of harness and memory, context management, and so on. I have my context managed in a structured (but nowadays way too big) Obsidian Vault. I also built myself a vector based \"vault search\" capability and have my harness use this as a tool to find thematically similar things across the dif", "context": "", "use_case": "", "switched_from": "", "story": "GPT-5.6"}]}, {"label": "Perceived degradation", "summary": "A few commenters feel quality dropped over time or after release.", "quotes": [{"id": 783097251, "item_id": 48943578, "url": "https://news.ycombinator.com/item?id=48943578", "by": "arman-w-jalili", "date": "2026-07-16", "stance": "warn", "quote": "i have the feeling GPT5.4 is not as good as before", "excerpt": "at the moment private i use mainly deepseek, about 80% and very very happy with it, at work we have copilot, i'm not sure, but the quality of service has got degraded in my opinion, speed is very low, i casually get error, api not available, especially gemeni-flash ,i have the feeling GPT5.4 is not as good as before, and with serious work i burn all my 400\u20ac budget in 4 days", "context": "", "use_case": "work LLM via Copilot", "switched_from": "", "story": "What % of your usage is open source vs. frontier models?"}, {"id": 794806673, "item_id": 49675417, "url": "https://news.ycombinator.com/item?id=49675417", "by": "asabla", "date": "2026-09-12", "stance": "warn", "quote": "I think this happened around Opus 4.5 or 4.5 and the same for GPT 5.4.", "excerpt": "> My new suspicion is now that they didn\u2019t got drastically smarter, but they got trained on the user input on the previous generations I think this happened around Opus 4.5 or 4.5 and the same for GPT 5.4. The more I use those models, harnesses, techniques for guidance etc etc. The more I land in going back to writing software by hand again. Maybe not all of it, but at least the crucial parts + foundations.", "context": "", "use_case": "coding", "switched_from": "", "story": "Anthropic boss Dario Amodei calls for AI development to slow down"}]}, {"label": "Knowledge cutoff", "summary": "Early GPT-5 struggled with newer tooling like uv and with R code.", "quotes": [{"id": 793919760, "item_id": 49619985, "url": "https://news.ycombinator.com/item?id=49619985", "by": "JLO64", "date": "2026-09-09", "stance": "warn", "quote": "it consistently ran the wrong commands or used pip despite my instructions within AGENTS.md", "excerpt": "Personally I can understand why these instructions were added. Back in the GPT-5 days I tried using that model for a uv/Python project but it consistently ran the wrong commands or used pip despite my instructions within AGENTS.md . This was because whatever uv features/behavior I was using was added after that model\u2019s cut off date. It got so bad that I switched to Claude which had a sooner knowledge cut-off. That sa", "context": "", "use_case": "uv/Python project coding", "switched_from": "", "story": "Codex silently begs agents to make arbitrary web requests"}, {"id": 789825217, "item_id": 49364076, "url": "https://news.ycombinator.com/item?id=49364076", "by": "5555watch", "date": "2026-08-19", "stance": "warn", "quote": "the coding with R language was terrible", "excerpt": ">Everyone should occasionally go back to the old models to see how much worse they were, like even a year ago you could generate results but they were typically full of bugs Oh yeah, I'm still amazed how good the current iteration of models are for coding (I have a fear it's too good to be true - so will get taken away..). Exactly a year ago I switched from GPT 5 to Gemini just because the coding with R language was ", "context": "", "use_case": "coding in R", "switched_from": "", "story": "Cerebras CS4"}]}], "who": [{"by": "nl", "context": "Heavy user for 3D model understanding since February", "url": "https://news.ycombinator.com/item?id=49930091", "date": "2026-10-02"}, {"by": "alexjplant", "context": "constrained by regulations at day job", "url": "https://news.ycombinator.com/item?id=49735497", "date": "2026-09-17"}, {"by": "ryptophan", "context": "Tech Lead at Georgian", "url": "https://news.ycombinator.com/item?id=49550977", "date": "2026-09-03"}, {"by": "alexcpn", "context": "experienced programmer heavily using AI for POCs", "url": "https://news.ycombinator.com/item?id=49373401", "date": "2026-08-20"}, {"by": "SwellJoe", "context": "building a harness for security work", "url": "https://news.ycombinator.com/item?id=49295164", "date": "2026-08-13"}, {"by": "runjake", "context": "on multiple $20 plans", "url": "https://news.ycombinator.com/item?id=49111328", "date": "2026-07-30"}, {"by": "guessmyname", "context": "builds harnesses for models", "url": "https://news.ycombinator.com/item?id=49045035", "date": "2026-07-24"}, {"by": "jdw64", "context": "on-site programmer", "url": "https://news.ycombinator.com/item?id=48912406", "date": "2026-07-14"}, {"by": "tda", "context": "not a C or Fortran dev", "url": "https://news.ycombinator.com/item?id=48907326", "date": "2026-07-14"}, {"by": "rh94", "context": "hiring engineers; codes entirely with AI", "url": "https://news.ycombinator.com/item?id=48857157", "date": "2026-07-10"}, {"by": "dalemhurley", "context": "Game developer since the late 90s", "url": "https://news.ycombinator.com/item?id=48829861", "date": "2026-07-08"}, {"by": "efficax", "context": "engineer on a team shipping product code", "url": "https://news.ycombinator.com/item?id=48796532", "date": "2026-07-05"}], "tags": ["for coding", "for agentic coding in Codex", "for code review", "for technical writing", "for reverse engineering"], "use_cases": ["coding", "ai coding", "agentic coding", "code review", "general coding", "coding agent"], "compared": [{"slug": "claude-opus", "name": "Claude Opus", "co": 104, "switched_from": 3, "switched_to": 0}, {"slug": "codex", "name": "Codex", "co": 54, "switched_from": 0, "switched_to": 0}, {"slug": "claude", "name": "Claude", "co": 30, "switched_from": 8, "switched_to": 1}, {"slug": "claude-fable", "name": "Claude Fable", "co": 39, "switched_from": 2, "switched_to": 1}, {"slug": "claude-sonnet", "name": "Claude Sonnet", "co": 23, "switched_from": 0, "switched_to": 2}, {"slug": "gemini", "name": "Gemini", "co": 24, "switched_from": 0, "switched_to": 1}, {"slug": "claude-code", "name": "Claude Code", "co": 22, "switched_from": 2, "switched_to": 0}, {"slug": "deepseek", "name": "DeepSeek", "co": 23, "switched_from": 0, "switched_to": 1}], "makers": [{"id": 781810864, "item_id": 48863179, "url": "https://news.ycombinator.com/item?id=48863179", "by": "tedsanders", "date": "2026-07-10", "stance": "recommend", "quote": "(I work at OpenAI.)", "excerpt": "> Even worse, it's not a fair comparison: they purposefully just used \"adaptive\" instead of \"max\" for Fable. We agree models should be compared on a fair basis. Unfortunately, adaptive was the only publicly available number. Anthropic doesn't generally let us run their models for evals, so we rely on whatever Anthropic or third parties have published. In this case, the Agents' Last Exam leaderboard has Fable Adaptive", "context": "works at OpenAI", "use_case": "AI model evals", "switched_from": "", "story": "GPT-5.6"}], "threads": [{"url": "https://news.ycombinator.com/item?id=48849066", "title": "GPT-5.6", "mentions": 16}, {"url": "https://news.ycombinator.com/item?id=48413629", "title": "Ask HN: What is your (AI) dev tech stack / workflow? (June 2026)", "mentions": 5}, {"url": "https://news.ycombinator.com/item?id=45904551", "title": "GPT-5.1: A smarter, more conversational ChatGPT", "mentions": 4}, {"url": "https://news.ycombinator.com/item?id=48835111", "title": "Grok 4.5", "mentions": 3}, {"url": "https://news.ycombinator.com/item?id=48689028", "title": "Previewing GPT\u20115.6 Sol: a next-generation model", "mentions": 3}, {"url": "https://news.ycombinator.com/item?id=48463808", "title": "Claude Fable 5", "mentions": 3}], "sentiment_by_month": {"2025-11": {"neutral": 5, "warn": 8, "recommend": 4}, "2025-12": {"warn": 3, "recommend": 4, "neutral": 2}, "2026-01": {"recommend": 6, "warn": 3, "neutral": 2}, "2026-02": {"neutral": 1, "warn": 2, "recommend": 2, "compare": 1}, "2026-03": {"neutral": 1, "recommend": 4}, "2026-04": {"recommend": 8, "neutral": 5, "warn": 3}, "2026-05": {"compare": 6, "recommend": 6, "neutral": 1}, "2026-06": {"recommend": 29, "warn": 8, "neutral": 13, "compare": 2}, "2026-07": {"recommend": 29, "compare": 14, "neutral": 10, "warn": 4}, "2026-08": {"recommend": 10, "neutral": 5, "warn": 2, "compare": 1}, "2026-09": {"recommend": 8, "neutral": 4, "compare": 2, "warn": 3}, "2026-10": {"recommend": 2, "compare": 1}}, "quotes": [{"id": 802761408, "item_id": 50172588, "url": "https://news.ycombinator.com/item?id=50172588", "by": "giancarlostoro", "date": "2026-10-20", "stance": "recommend", "quote": "GPT 5.5 on Medium is all I use lately, I don't understand why anyone would pay drastically more for slightly more", "excerpt": "GPT 5.5 on Medium is all I use lately, I don't understand why anyone would pay drastically more for slightly more, when Opus has been good enough for a year now to do most things.", "context": "", "use_case": "general coding", "switched_from": "", "story": "SlopCodeBench 10/2026 \u2013 Opus 5.5, Astra wins overall, xhigh considered harmful"}, {"id": 799675651, "item_id": 49979728, "url": "https://news.ycombinator.com/item?id=49979728", "by": "piscaries", "date": "2026-10-06", "stance": "compare", "quote": "Once on Codex with GPT-5.6.", "excerpt": "Author here. The original article is here: https://piscaries.github.io/posts/codex-and-pi-build-the-same-app-in-orca/ I had two goals: 1. try out Orca for running a team of coding agents 2. See how Pi, a minimal open-source agent, generates codes against a popular one like Codex. So I ran the same small team (spec, design, build, with a shared reviewer) twice in Orca. Once on Codex with GPT-5.6. The other on Pi with ", "context": "", "use_case": "model powering Codex", "switched_from": "", "story": "I had Codex and Pi build the same app in Orca, then compared them side by side"}, {"id": 798881456, "item_id": 49930091, "url": "https://news.ycombinator.com/item?id=49930091", "by": "nl", "date": "2026-10-02", "stance": "recommend", "quote": "Opus 4.8 was usable, but the GPT 5.x series was better.", "excerpt": "Having used them heavily for 3D model understanding since February I can say it's nuanced. Opus 4.6 and 4.7 were bad, but GPT 5.2 and above were very usable. Opus 4.8 was usable, but the GPT 5.x series was better. Fable is great. Opus 5.0 was interesting. It could solve some problems that Sol 5.x couldn't solve (applying a G2 curve on a 3 way corner where one face was a Bezier curve) but you had to be super prescript", "context": "Heavy user for 3D model understanding since February", "use_case": "3D model understanding", "switched_from": "", "story": "Using Opus 5.5 to discover a new eyewitness record of the dodo"}, {"id": 797751586, "item_id": 49859474, "url": "https://news.ycombinator.com/item?id=49859474", "by": "tptacek", "date": "2026-09-26", "stance": "recommend", "quote": "GPT4 and GPT5 were more than enough supplementation", "excerpt": "I'm a broken record on how valuable I found Math Academy to be; 22,000 XP in roughly a year, going from high school trig (and a grim confrontation with my intuitive ability to work with fractions) through multivariable calculus. The most important thing to understand about Math Academy is that it's all problems, problems over and over again, with just enough pedagogy to get you from one kind of problem to the next. I", "context": "", "use_case": "deeper understanding of math topics", "switched_from": "", "story": "Reflections on 1,000 Days of Math"}, {"id": 797664129, "item_id": 49854008, "url": "https://news.ycombinator.com/item?id=49854008", "by": "epihelix", "date": "2026-09-26", "stance": "recommend", "quote": "It writes very nice, clean code with almost no comments.", "excerpt": "Yes, with Enterprise M365 Copilot. Select \"GPT 5.6 Think deeper\" (ugh, so unrevealing, MS) and you've got free, unlimited Sol or Terra (likely a mix depending on load) on tap. It writes very nice, clean code with almost no comments. I use it a lot.", "context": "", "use_case": "writing code", "switched_from": "", "story": "Microsoft abandons personal AI chatbot race with Copilot reboot"}, {"id": 796639233, "item_id": 49789952, "url": "https://news.ycombinator.com/item?id=49789952", "by": "jotato", "date": "2026-09-21", "stance": "neutral", "quote": "It was just as good as 5.5 which was my previous default.", "excerpt": "Just yesterday I was thinking about gpt-5.6-luna. I made it my default model in Hermes during its fist week of launch. It was just as good as 5.5 which was my previous default. But over the last 2 or 3 weeks I've seen how dumb it is now. I have to be very explicit with it. For example, I used to be able to prompt \"Check the system logs on <server> for....\" and it would just figure it out. Yesterday I asked \"Did <serv", "context": "", "use_case": "default model in Hermes", "switched_from": "", "story": "Fable 5 \u2013 Median thinking declined in August"}, {"id": 796072609, "item_id": 49754538, "url": "https://news.ycombinator.com/item?id=49754538", "by": "matt_oriordan", "date": "2026-09-18", "stance": "compare", "quote": "GPT-5.6 Sol 3.5 s", "excerpt": "Hi HN, Matt here, co-founder of Ably. I built this after Jev landed on Vercel's AI Gateway: https://github.com/ably-labs/jev-pong It's pong where the ball moves one step per model decision. Four lanes, one model each, same serve, same rules. Every step the model gets the same tiny JSON state (ball, paddle, predicted intercept) and the same question: up, down or stay. One answer moves the ball one segment; eight segme", "context": "", "use_case": "Pong decision latency", "switched_from": "", "story": "Show HN: Jev vs. GPT-5.6 and Claude Haiku at Pong"}, {"id": 795767955, "item_id": 49735497, "url": "https://news.ycombinator.com/item?id=49735497", "by": "alexjplant", "date": "2026-09-17", "stance": "recommend", "quote": "My recent experiences with GPT-5.6 were also very positive in this regard.", "excerpt": "I bother because Claude Code does wacky nonsense unless I use OpenSpec (or something similar) to explicitly research, scope, persist, then implement in a brand-new context. Even Opus will do ridiculous nonsense like throw its hands up when asked to start a Docker container, ignore explicit architectural instructions, write verbose make-work documentation riddled with inaccuracies, etc. Using OpenSpec keeps things as ", "context": "constrained by regulations at day job", "use_case": "AI coding", "switched_from": "", "story": "OpenSpec \u2013 A lightweight and configurable AI spec framework"}, {"id": 795703408, "item_id": 49731463, "url": "https://news.ycombinator.com/item?id=49731463", "by": "MAZ97", "date": "2026-09-16", "stance": "recommend", "quote": "Gpt 5.6, it understands me perfectly and is able to always do what i tell it to the point", "excerpt": "Gpt 5.6, it understands me perfectly and is able to always do what i tell it to the point", "context": "", "use_case": "", "switched_from": "", "story": "Ask HN: What default model do you use and why?"}, {"id": 795687536, "item_id": 49730471, "url": "https://news.ycombinator.com/item?id=49730471", "by": "amluto", "date": "2026-09-16", "stance": "neutral", "quote": "I used GPT-5.5, IIRC, which was capable of doing the math I needed correctly, but only after a lot of tries", "excerpt": "Did any of the prompts involve the user trying it and observing that the scrubber barely works and that the topography texturing (the displacement map or normal map or whatever technique GPT-6 picked) is outrageously wrong? I'm no expert in tectonic movements, but I guarantee that continental plates don't slide around without the mountains on them moving along with them. It is shiny, though. (I have personally once v", "context": "", "use_case": "vibecoding a 3D viewer", "switched_from": "", "story": "GPT-6 built this earth exploration site in 5 prompts"}, {"id": 794806673, "item_id": 49675417, "url": "https://news.ycombinator.com/item?id=49675417", "by": "asabla", "date": "2026-09-12", "stance": "warn", "quote": "I think this happened around Opus 4.5 or 4.5 and the same for GPT 5.4.", "excerpt": "> My new suspicion is now that they didn\u2019t got drastically smarter, but they got trained on the user input on the previous generations I think this happened around Opus 4.5 or 4.5 and the same for GPT 5.4. The more I use those models, harnesses, techniques for guidance etc etc. The more I land in going back to writing software by hand again. Maybe not all of it, but at least the crucial parts + foundations.", "context": "", "use_case": "coding", "switched_from": "", "story": "Anthropic boss Dario Amodei calls for AI development to slow down"}, {"id": 794782417, "item_id": 49673901, "url": "https://news.ycombinator.com/item?id=49673901", "by": "gpugreg", "date": "2026-09-12", "stance": "warn", "quote": "they often give up saying it can't be done, or cheat by modifying the tests", "excerpt": "I currently use DeepSeek-V4.1-Flash and the V4-Pro and -Flash versions before that, because they are extremely cheap. I have spent less than $25 for over a billion token so far (1B cached, 15M out, 19M in). I even prefer the DeepSeek models over the older OpenAI offerings. When I tell a GPT-5.x model to do some difficult task, they often give up saying it can't be done, or cheat by modifying the tests, while the Deep", "context": "", "use_case": "difficult tasks", "switched_from": "", "story": "Ask HN: What default model do you use and why?"}, {"id": 794421985, "item_id": 49651374, "url": "https://news.ycombinator.com/item?id=49651374", "by": "alex-moon", "date": "2026-09-10", "stance": "recommend", "quote": "At work I use Codex CLI with GPT 5.4 - it hits a sweet spot", "excerpt": "Benchmaxxing is the word you're reaching for here (though code golf is a perfect name for the symptom). At work I use Codex CLI with GPT 5.4 - it hits a sweet spot: capable enough that it'll do the right thing most of the time, but unsophisticated enough that I can follow along with what it's doing and cut it off and go, \"Nope! That's wrong because x, try y instead.\" For users like me, the threshold for PMF was never", "context": "", "use_case": "coding at work", "switched_from": "", "story": "Astra for Coding: Why Are We Doing This Again?"}, {"id": 793919760, "item_id": 49619985, "url": "https://news.ycombinator.com/item?id=49619985", "by": "JLO64", "date": "2026-09-09", "stance": "warn", "quote": "it consistently ran the wrong commands or used pip despite my instructions within AGENTS.md", "excerpt": "Personally I can understand why these instructions were added. Back in the GPT-5 days I tried using that model for a uv/Python project but it consistently ran the wrong commands or used pip despite my instructions within AGENTS.md . This was because whatever uv features/behavior I was using was added after that model\u2019s cut off date. It got so bad that I switched to Claude which had a sooner knowledge cut-off. That sa", "context": "", "use_case": "uv/Python project coding", "switched_from": "", "story": "Codex silently begs agents to make arbitrary web requests"}, {"id": 793433441, "item_id": 49589590, "url": "https://news.ycombinator.com/item?id=49589590", "by": "senko", "date": "2026-09-06", "stance": "compare", "quote": "Compare Astra[0] with GPT 5.4[1] which was OpenAI's state of the art just six months ago.", "excerpt": "I use Minecraft (and Warcraft and an arcade flying simulator) as a silly but directionally correct indication of the models' capabilitites. Compare Astra[0] with GPT 5.4[1] which was OpenAI's state of the art just six months ago. (all tests on more models with code and prompts available here: https://senko.net/vibecode-bench ) Yes, it's not a scientific benchmark but it's a good heuristic. For a better eval, create a", "context": "", "use_case": "vibe-coding benchmark", "switched_from": "", "story": "Recreating Minecraft Is Not a Benchmark"}, {"id": 793242849, "item_id": 49577678, "url": "https://news.ycombinator.com/item?id=49577678", "by": "Topfi", "date": "2026-09-05", "stance": "recommend", "quote": "GPT-5.4 (which remains the best model by any lab I ever tested for task adherence", "excerpt": "Far too early for any real opinion, but GPT-6 Astra (on Light) stops a lot and often in unintuitive ways that I haven't seen in a while. Had a very hard to reproduce windows focus bug on macOS that was unreliable to reproduce and when it could be reproduced, it disappeared the second I switched to the app formerly known as Codex, so for the first time used the voice interface with that little hovering bubble, telling", "context": "", "use_case": "task adherence", "switched_from": "", "story": "Ask HN: Initial Thoughts on GPT-6 Astra?"}, {"id": 793242850, "item_id": 49577678, "url": "https://news.ycombinator.com/item?id=49577678", "by": "Topfi", "date": "2026-09-05", "stance": "warn", "quote": "GPT-5.5 simply did not stay on track post compaction", "excerpt": "Far too early for any real opinion, but GPT-6 Astra (on Light) stops a lot and often in unintuitive ways that I haven't seen in a while. Had a very hard to reproduce windows focus bug on macOS that was unreliable to reproduce and when it could be reproduced, it disappeared the second I switched to the app formerly known as Codex, so for the first time used the voice interface with that little hovering bubble, telling", "context": "", "use_case": "", "switched_from": "", "story": "Ask HN: Initial Thoughts on GPT-6 Astra?"}, {"id": 792976513, "item_id": 49561032, "url": "https://news.ycombinator.com/item?id=49561032", "by": "the_duke", "date": "2026-09-04", "stance": "neutral", "quote": "With the skill installed GPT 5.6 usually automatically uses it", "excerpt": "I've had great success with a tool I wrote that can print sparse ASTs for code. https://github.com/theduke/smartedit With the skill installed GPT 5.6 usually automatically uses it, and it reduces code exploration time and token usage significantly, for example by just printing the types and functions in a file without bodies, and only expanding when needed. (note: it also has editing functionality, which doesn't work", "context": "", "use_case": "coding agent", "switched_from": "", "story": "Grep beats LSP? Why coding agents ignore your fancier tools"}, {"id": 792877042, "item_id": 49554815, "url": "https://news.ycombinator.com/item?id=49554815", "by": "wahnfrieden", "date": "2026-09-03", "stance": "recommend", "quote": "whereas I was able to work ~80 hours/week with 5.5 High", "excerpt": "2.5x more expensive than Sol. Can expect 2.5x more usage in Codex subscription. Sol is already brutal (even after their recent fixes, it's just a token-hungry model: I go through a full 20x account per day, on Sol Med/High standard speed, with ~2 threads). Note that Tibo recommended using Sol Med as daily driver. When I'm doing less complicated work, I can't even make it past 2-3 days with Sol Med, whereas I was able", "context": "", "use_case": "coding", "switched_from": "", "story": "OpenAI begins rolling out GPT-6 Astra"}, {"id": 792815634, "item_id": 49550977, "url": "https://news.ycombinator.com/item?id=49550977", "by": "ryptophan", "date": "2026-09-03", "stance": "neutral", "quote": "gemini-2.5-pro, claude-opus-4.8, gpt-5.1, llama-3.3-70b", "excerpt": "Author here. I'm a Tech Lead at Georgian (a growth-stage VC). Short version: I used a few retrieval benchmark datasets, and experimented with a 21 embedding model leaderboard. I held the benchmark corpus and relevance labels fixed, and transformed the queries: paraphrase, terse, verbose. I kept the \u201cinformation need\u201d the same under each query transformation, but the leaderboard order changed more than I expected. I r", "context": "Tech Lead at Georgian", "use_case": "rewriting benchmark queries", "switched_from": "", "story": "Your Queries Can Reorder the Retrieval Leaderboard"}, {"id": 792321072, "item_id": 49520067, "url": "https://news.ycombinator.com/item?id=49520067", "by": "ramon156", "date": "2026-09-01", "stance": "recommend", "quote": "the usage and quality out of gpt 5.6 is on par, if not better in token usage.", "excerpt": "the usage and quality out of gpt 5.6 is on par, if not better in token usage. I would love to do a write-up on which tools are useful for what use-case, as AA keeps improving. I use gemini for rewriting code docs, because the frontier models are so verbose when it comes to writing text", "context": "", "use_case": "coding", "switched_from": "", "story": "Is it just me, or has Claude Opus gotten worse recently?"}, {"id": 791913570, "item_id": 49494598, "url": "https://news.ycombinator.com/item?id=49494598", "by": "thefourthchime", "date": "2026-08-30", "stance": "neutral", "quote": "about 10x faster than just using Opus 5 or GPT 5.6 and about 90% is good", "excerpt": "It's a little confusing because Grok and Cursor kinda live in different worlds. Grok has SuperGrok and SuperGrok Heavy ($300 a month!) The best bet is Cursor Ultra ($200), which is the equivalent a month plan that you get from OpenAI and Anthropic. I get a LOT more capacity out of it because I almost always use the \"Auto\" mode, which auto picks between Opus 5, GPT 5.6, Grok 4.6, and Composer 2.5. I like Auto mode so ", "context": "", "use_case": "coding", "switched_from": "", "story": "Our decision on Cursor following its acquisition by SpaceX"}, {"id": 791494016, "item_id": 49468376, "url": "https://news.ycombinator.com/item?id=49468376", "by": "FooBarWidget", "date": "2026-08-27", "stance": "recommend", "quote": "With GPT-5.6 I get reasonable results with \"use natural, plain English\".", "excerpt": "I think this is more Claude writing than LLM writing. GPT-5.6 also talks like an LLM, but not like this. With GPT-5.6 I get reasonable results with \"use natural, plain English\". I tried \"use ASD-STE100 Simplified Technical English\" but I've found \"natural plain English\" to work better. Still not on the level I would write, but better. In the mean time, I hear that Claude resists changing its writing style. A PO used ", "context": "", "use_case": "writing in plain English", "switched_from": "", "story": "Show HN: Restoredrill \u2013 proves your Postgres backups restore"}, {"id": 791043313, "item_id": 49440207, "url": "https://news.ycombinator.com/item?id=49440207", "by": "zerof1l", "date": "2026-08-25", "stance": "recommend", "quote": "Finally switched to GPT 5.6 Sol and Terra, and it's so much better.", "excerpt": "The issue with models and Claude in general is that I don\u2019t know anymore what model selection and thinking level are for. When first released, Opus 5 worked well; now it's crap. We have no control or visibility over what is really happening on the other end. You don\u2019t know what is really running on the other end. Our incentives are misaligned: Anthropic gets paid per token - some abstract unit of measure that keeps c", "context": "", "use_case": "coding", "switched_from": "Claude Opus 5", "story": "Anthropic Sees over $30T in Potential Revenue"}, {"id": 789974416, "item_id": 49373401, "url": "https://news.ycombinator.com/item?id=49373401", "by": "alexcpn", "date": "2026-08-20", "stance": "recommend", "quote": "I find 5.5 models more than enough for most tasks and have shifted back to codex", "excerpt": "Really strange; anyway it is closed source and best not to have too much reliance and attachment to any one tool. I have switched to GPT 5.5 and 5.6 Sol and Codex ove Claude already. Fable is very expensive/token limiting. I am an experienced programmer heavily using AI for many POCs now. At one time Anthropic and its models Opus,Fable looked leading; now OpenAI has caught up. I find 5.5 models more than enough for m", "context": "experienced programmer heavily using AI for POCs", "use_case": "programming POCs", "switched_from": "Claude", "story": "Anthropic Refuses to Support Agents.md"}, {"id": 789825217, "item_id": 49364076, "url": "https://news.ycombinator.com/item?id=49364076", "by": "5555watch", "date": "2026-08-19", "stance": "warn", "quote": "the coding with R language was terrible", "excerpt": ">Everyone should occasionally go back to the old models to see how much worse they were, like even a year ago you could generate results but they were typically full of bugs Oh yeah, I'm still amazed how good the current iteration of models are for coding (I have a fear it's too good to be true - so will get taken away..). Exactly a year ago I switched from GPT 5 to Gemini just because the coding with R language was ", "context": "", "use_case": "coding in R", "switched_from": "", "story": "Cerebras CS4"}, {"id": 789512418, "item_id": 49344526, "url": "https://news.ycombinator.com/item?id=49344526", "by": "Aperocky", "date": "2026-08-18", "stance": "recommend", "quote": "The communication is more direct too with GPT5.6 family of LLM.", "excerpt": "One of the biggest reason I migrated to Codex. I'd like to see each agentic CLI process as a simple building block, and having 100 claude session means I sometimes run out of memory (and I have 256GB of them). Codex never have this problem. The communication is more direct too with GPT5.6 family of LLM. No more belt and suspenders.", "context": "", "use_case": "coding", "switched_from": "", "story": "Linux 7.3 improves performance when running out of vRAM"}, {"id": 789504304, "item_id": 49344019, "url": "https://news.ycombinator.com/item?id=49344019", "by": "justinbaker84", "date": "2026-08-18", "stance": "recommend", "quote": "I switched from claude to gpt when 5.6 came out for the same reasons.", "excerpt": "I switched from claude to gpt when 5.6 came out for the same reasons. I don't understand why so many people are still using claude when GPT and open source are so much better.", "context": "", "use_case": "AI assistant / coding", "switched_from": "Claude", "story": "GPT-5.6 Sol Pricing Cut by 50%"}, {"id": 789209377, "item_id": 49325586, "url": "https://news.ycombinator.com/item?id=49325586", "by": "SwellJoe", "date": "2026-08-16", "stance": "recommend", "quote": "GPT 5.5 did a similar task in about 20 minutes.", "excerpt": "This is true, but I think it understates the problem. I did a task I've done with a bunch of small models lately (https://github.com/swelljoe/flar/pull/17), and it did an excellent job, the best of any self-hostable model. But, it took eleven (11!) hours on my dual GPU setup. It really chewed on it, and spent a lot of time checking and re-checking. It is by far the slowest model I've used for the task. GPT 5.5 did a ", "context": "", "use_case": "coding task", "switched_from": "", "story": "Qwen 3.8 27B is excellent, but it defaults to overthinking things"}, {"id": 789098242, "item_id": 49318640, "url": "https://news.ycombinator.com/item?id=49318640", "by": "simonw", "date": "2026-08-16", "stance": "recommend", "quote": "I've not caught it with the more recent (Opus 4.5+, Fable 5, GOT 5.5/5.6) models.", "excerpt": "Have you seen that recently? I used to see that happen, but I've not caught it with the more recent (Opus 4.5+, Fable 5, GOT 5.5/5.6) models.", "context": "", "use_case": "coding", "switched_from": "", "story": "Software Engineering fundamentals matter more"}, {"id": 788852401, "item_id": 49303275, "url": "https://news.ycombinator.com/item?id=49303275", "by": "jmstfv", "date": "2026-08-14", "stance": "recommend", "quote": "Both GPT 5.5 and 5.6 are quite capable", "excerpt": "I've switched to Codex a few months ago when Claude's weekly limits were getting pretty stiff, and I haven't looked back. Both GPT 5.5 and 5.6 are quite capable, especially compared to nerfed Opus 4.7 (haven't tried 5). Also, the Codex guy regularly resets weekly limits for everyone, which is a nice bonus (I know it's a temporary gimmick to attract more users, but I might as well use it while it lasts.)", "context": "", "use_case": "AI coding", "switched_from": "", "story": "Why does Opus 5 feel worse to work with?"}, {"id": 788852402, "item_id": 49303275, "url": "https://news.ycombinator.com/item?id=49303275", "by": "jmstfv", "date": "2026-08-14", "stance": "recommend", "quote": "Both GPT 5.5 and 5.6 are quite capable", "excerpt": "I've switched to Codex a few months ago when Claude's weekly limits were getting pretty stiff, and I haven't looked back. Both GPT 5.5 and 5.6 are quite capable, especially compared to nerfed Opus 4.7 (haven't tried 5). Also, the Codex guy regularly resets weekly limits for everyone, which is a nice bonus (I know it's a temporary gimmick to attract more users, but I might as well use it while it lasts.)", "context": "", "use_case": "AI coding", "switched_from": "", "story": "Why does Opus 5 feel worse to work with?"}, {"id": 788722625, "item_id": 49295164, "url": "https://news.ycombinator.com/item?id=49295164", "by": "SwellJoe", "date": "2026-08-13", "stance": "neutral", "quote": "had to switch to GPT 5.5 when even Opus started refusing security work", "excerpt": "I've been building a harness for security work, and had to switch to GPT 5.5 when even Opus started refusing security work. Then 5.6 Sol arrived, and it refuses security work, too. So, I switched to Kimi K3 and DeepSeek for API testing just because it's so much cheaper. But, if GLM is better, I'm here for it, as I think GLM is also cheaper than K3.", "context": "building a harness for security work", "use_case": "security work", "switched_from": "Opus", "story": "GLM-5.3: Frontier Coding with Emergent Cyber Capabilities"}, {"id": 788680099, "item_id": 49292506, "url": "https://news.ycombinator.com/item?id=49292506", "by": "gazebo2", "date": "2026-08-13", "stance": "compare", "quote": "it's definitely powerful but I honestly think I'll go back to Gemini", "excerpt": "I similarly have a weird affinity for Gemini that I can't really articulate. I used Gemini's free chat and found it great for exploring technical topics (and random one-off general walking-around-questions) and appreciated its speed, tone and accuracy. I spent a month playing with Gemini CLI / Antigravity and found it also an effective coding agent, at least for my workflow (entirely in the loop development and revie", "context": "", "use_case": "coding", "switched_from": "", "story": "Gemini 3.7 Flash"}, {"id": 788237058, "item_id": 49264816, "url": "https://news.ycombinator.com/item?id=49264816", "by": "simonw", "date": "2026-08-11", "stance": "neutral", "quote": "Looks like it was running Claude Opus 4.6 and GPT-5.4.", "excerpt": "This is slightly old news at this point - the BSides conference was in March, this article was published in May, we've had a whole lot of additional evidence since May that makes it unsurprising that AI can solve CTF challenges! Here's the repo mentioned in the story: https://github.com/verialabs/ctf-agent > Autonomous CTF solver that races multiple AI models in parallel. 1st place BSidesSF 2026. That one hasn't had ", "context": "", "use_case": "CTF solving", "switched_from": "", "story": "AI Is Solving CTF Challenges in Minutes"}, {"id": 788015744, "item_id": 49250984, "url": "https://news.ycombinator.com/item?id=49250984", "by": "ferris-booler", "date": "2026-08-10", "stance": "recommend", "quote": "GPT-5.6 so far has been a significantly better technical writer in my opinion, and it's much faster in conversations.", "excerpt": "Yeah I hit this feeling with Claude one too many times and switched from Anthropic Pro to OpenAI Pro. GPT-5.6 so far has been a significantly better technical writer in my opinion, and it's much faster in conversations. My guess is that Claude's fantasy-jargon is a symptom of training failure, not a sneaky intelligence edge (I could be wrong). Tangential but I use OpenCode with GPT-5.6 rather than Codex because I cou", "context": "", "use_case": "technical writing and conversation", "switched_from": "Claude", "story": "Humanising LLM Outputs Is Dumb"}, {"id": 787860976, "item_id": 49241311, "url": "https://news.ycombinator.com/item?id=49241311", "by": "vincent_s", "date": "2026-08-10", "stance": "warn", "quote": "It definitely feels like subscription limits have dropped since GPT-5.6 came out.", "excerpt": "It definitely feels like subscription limits have dropped since GPT-5.6 came out. But I couldn't really find any evidence for it, going through my history. All I could come up with is that 5.6-sol uses about twice the tokens compared to 5.5 (both on xhigh) [0]. One thing I just did was to stop four long-running Codex sessions that ran on 5.6-sol and switched to 5.5 and asked it \"Please check if you're really going ag", "context": "", "use_case": "coding", "switched_from": "", "story": "Switching from GPT-5.5 to GPT-5.6 Made Me Less Productive"}, {"id": 787860977, "item_id": 49241311, "url": "https://news.ycombinator.com/item?id=49241311", "by": "vincent_s", "date": "2026-08-10", "stance": "recommend", "quote": "stop four long-running Codex sessions that ran on 5.6-sol and switched to 5.5", "excerpt": "It definitely feels like subscription limits have dropped since GPT-5.6 came out. But I couldn't really find any evidence for it, going through my history. All I could come up with is that 5.6-sol uses about twice the tokens compared to 5.5 (both on xhigh) [0]. One thing I just did was to stop four long-running Codex sessions that ran on 5.6-sol and switched to 5.5 and asked it \"Please check if you're really going ag", "context": "", "use_case": "coding", "switched_from": "GPT-5.6", "story": "Switching from GPT-5.5 to GPT-5.6 Made Me Less Productive"}, {"id": 787474371, "item_id": 49217148, "url": "https://news.ycombinator.com/item?id=49217148", "by": "dools", "date": "2026-08-07", "stance": "neutral", "quote": "if deepseek can\u2019t figure something out after a few goes I switch to GTP 5.5 or 5.6", "excerpt": "I have been using deepseek v4 pro almost exclusively. I was using Kimi a lot but it just nose dived. The decline started with the release of 2.7 and accelerated with the release of 3. When I need vision capabilities I use GPT 5.3 codex and if deepseek can\u2019t figure something out after a few goes I switch to GTP 5.5 or 5.6 (I\u2019ve been giving Terra first bite recently and it does pretty well, and have used Sol a couple o", "context": "", "use_case": "fallback for hard problems", "switched_from": "", "story": "DeepSeek V4 Flash 0731"}, {"id": 787223539, "item_id": 49201471, "url": "https://news.ycombinator.com/item?id=49201471", "by": "superfrank", "date": "2026-08-06", "stance": "recommend", "quote": "I've been using this with GPT 5.5 and now 5.6 for a few months and running it a few times a week with no issue.", "excerpt": "I've come to the same conclusion over and over with all of the Chinese models that have been claimed to be catching up with OpenAI's and Anthropic's frontier models (Deepseek 4, GLM 5.2, Kimi K3). At their best, I think they're closing in on Opus and GPT, but they're incredibly inconsistent and the variance in output quality is much higher than the best from any of the Anthropic or OpenAI models from the last few gen", "context": "", "use_case": "data sanity checks", "switched_from": "", "story": "Qwen3.8 Max now ranked as the best overall model by agentic index"}, {"id": 786465761, "item_id": 49154110, "url": "https://news.ycombinator.com/item?id=49154110", "by": "sdoering", "date": "2026-08-03", "stance": "recommend", "quote": "switching back to gpt-5.5 for \"real work\"", "excerpt": "In my daily work, I found expeciall terra and sol now stopping every few rounds again, telling me the tak is done. I even had them create a detailled plan - and told them to finish \"end to end\" - and they appruptly stop after the plan. Because they interpret this as finished. Even if the DOD is clearly not \"finish the plan\". The new models are shite (pardon my French), when it comes to long running tasks and I find m", "context": "", "use_case": "real work", "switched_from": "GPT-5.6 Sol", "story": "The Shape of Things to Come"}, {"id": 785781250, "item_id": 49111328, "url": "https://news.ycombinator.com/item?id=49111328", "by": "runjake", "date": "2026-07-30", "stance": "warn", "quote": "5.6 seems to use too many tokens to do the same or slightly inferior job", "excerpt": "Sonnet 5 Medium and GPT 5.5 Medium for coding. I drop down to a Mini model for file-based tasks. Codex keeps trying to revert me back to 5.6 but 5.6 seems to use too many tokens to do the same or slightly inferior job. I am only on multiple $20 plans.", "context": "on multiple $20 plans", "use_case": "coding", "switched_from": "", "story": "Ask HN: Which one do you use for planning and coding between sonnet and Opus?"}, {"id": 784010610, "item_id": 49000663, "url": "https://news.ycombinator.com/item?id=49000663", "by": "caspar", "date": "2026-07-21", "stance": "recommend", "quote": "GPT 5.4/5.5 would prefer them.", "excerpt": "This is partially a \"how much does the model follow instructions\" thing. I use Pi with a vibed LSP extension and Claude (4.6 or thereabouts) almost never followed instructions to use LSP renaming tools - despite it being strongly emphasized in system prompt and agents.md. However I found Codex 5.3 would use them sometimes, and GPT 5.4/5.5 would prefer them.", "context": "", "use_case": "LSP refactoring", "switched_from": "", "story": "Stop Using OpenCode"}, {"id": 781862193, "item_id": 48866387, "url": "https://news.ycombinator.com/item?id=48866387", "by": "arendtio", "date": "2026-07-10", "stance": "recommend", "quote": "the GPT 5.6 family does a good job in terms of cost for performance", "excerpt": "I also like to take a look at https://cursor.com/cursorbench While in the past months Composer 2.5 was a lot better than I had expected a year ago, and the GPT 5.6 family does a good job in terms of cost for performance, I wonder why nobody is talking about Grok 4.5 high? Those numbers look very convincing to me.", "context": "", "use_case": "coding", "switched_from": "", "story": "GPT-5.6"}, {"id": 781598688, "item_id": 48849918, "url": "https://news.ycombinator.com/item?id=48849918", "by": "edumucelli", "date": "2026-07-09", "stance": "recommend", "quote": "GPT won hands down.", "excerpt": "Not sure about the consensus, but during an entire week I have done every task on my workplace with both Opus 4.8 and GPT 5.5. GPT won hands down. I would even sometimes copy the plans and solutions (using different Git worktrees) from GPT and paste it on Opus and itself would say GPT plans were better. At that point I have migrated. Fable is not enabled in our workspace so I have not tried. Claude lost my trust arou", "context": "", "use_case": "workplace coding tasks", "switched_from": "Claude", "story": "GPT-5.6"}, {"id": 781591552, "item_id": 48849472, "url": "https://news.ycombinator.com/item?id=48849472", "by": "jakswa", "date": "2026-07-09", "stance": "recommend", "quote": "With gpt 5.5 I don't think I'm working differently but haven't felt the need to check anything because I think I've hit my limit... once?", "excerpt": "I'll agree and expand on \"weird restrictions\" -- I used to check the claude usage graphs multiple times a day to see where I'm at on my weekly budget. With gpt 5.5 I don't think I'm working differently but haven't felt the need to check anything because I think I've hit my limit... once? on some egregious edge case scenario iirc", "context": "", "use_case": "coding", "switched_from": "Claude", "story": "GPT-5.6"}]}