{"slug": "gpt-4", "name": "GPT-4", "category": "ai-models", "category_name": "AI models", "description": "", "recommenders": 33, "commenters": 119, "warners": 23, "score": 21.932, "rising": -6.82, "first": "2025-11-02", "latest": "2026-10-19", "why": "GPT\u20114o wasn\u2019t just a tool; it was my creative partner.", "why_is_quote": true, "has_verdict": false, "sentiment": {"recommend": 33, "neutral": 53, "compare": 15, "warn": 23}, "spark": [3, 3, 6, 13, 3, 0, 1, 1, 0, 1, 1, 1], "listed": true, "loved": [], "caveat": "", "short_quote": "GPT\u20114o wasn\u2019t just a tool; it was my creative partner.", "kind": "", "icon": true, "family": "gpt", "domain": "openai.com", "github_repo": null, "official_url": "https://openai.com/index/gpt-4/", "pricing": "", "aliases": ["GPT-4o", "GPT-4", "GPT-4.5", "GPT-4.1", "GPT-4o-mini", "GPT-4o mini", "gpt-4o-transcribe", "GPT-4.1 mini", "gpt-4chan", "GPT-4o Transcribe", "GPT-4 Vision", "GPT-4 Turbo", "GPT-4.1-mini", "GPT-4o Realtime"], "verdict": null, "praise": [{"label": "", "summary": "", "quotes": [{"id": 802513056, "item_id": 50157066, "url": "https://news.ycombinator.com/item?id=50157066", "by": "minraws", "date": "2026-10-19", "version": "4", "stance": "recommend", "quote": "GPT-3 vs GPT-4 was practically a 10-100x jump", "excerpt": "It's not really if you mean to point to the bitter lesson, it was models/intelligence scales better with general/simple methods with more compute better than it scales with any hand-crafted human program/method. Ala intelligence is a function of compute, there was no guarantee that you need to 10x compute for 10% better performance or something like that. If you ask the labs they will say the scaling is fairly linear", "context": "", "use_case": "", "switched_from": "", "story": "MiniMax M3.1 Flash"}, {"id": 797751585, "item_id": 49859474, "url": "https://news.ycombinator.com/item?id=49859474", "by": "tptacek", "date": "2026-09-26", "version": "4", "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": 792198881, "item_id": 49512430, "url": "https://news.ycombinator.com/item?id=49512430", "by": "piltdownman", "date": "2026-08-31", "version": "4", "stance": "recommend", "quote": "This study suggests that GPT-4 models could be leveraged as a support tool in mental health telephone triage", "excerpt": "Well, yes. Not a huge breadth of research, but all conclusions support my statements above. Two such relevant and recent studies: Evaluation of an Artificial Intelligence and Online Psychotherapy Initiative to Improve Access and Efficiency in an Ambulatory Psychiatric Setting https://pmc.ncbi.nlm.nih.gov/articles/PMC12316677/ //The overall wait time to receive care decreased by 71.43% due to this initiative. Addition", "context": "", "use_case": "", "switched_from": "", "story": "Mental Health Workers Say Algorithmic Triage Is Hurting Patients"}, {"id": 774718432, "item_id": 48419902, "url": "https://news.ycombinator.com/item?id=48419902", "by": "stared", "date": "2026-06-05", "version": "4", "stance": "recommend", "quote": "GPT4, when it could do a translation that would take a considerable human effort", "excerpt": "GPT4, when it could do a translation that would take a considerable human effort, vide \"Genesis 1 but every word begins with 'A'\": https://p.migdal.pl/blog/2023/05/genesis-az-by-gpt/", "context": "", "use_case": "translation", "switched_from": "", "story": "What was your \"Oh Shit\" moment with GenAI?"}, {"id": 772601106, "item_id": 48287569, "url": "https://news.ycombinator.com/item?id=48287569", "by": "carpo", "date": "2026-05-26", "version": "4.1", "stance": "recommend", "quote": "the best one so far is GPT4.1-Mini", "excerpt": "Yeah, it's been awesome! I'm so excited about tool calls and function use, the possibilities are huge. I ran it over 1494 videos that range in length from a few seconds to over 3 hours. Total duration 260 hours and a total size of 3795 GB. I don't know exactly how long it took to run, as I found some bugs I needed to fix when processing mkv files, but it was probably around 24 hours in total. That wasn't all LLM requ", "context": "", "use_case": "RAG chat over video library", "switched_from": "", "story": "Indexing a year of video locally on a 2021 MacBook with Gemma4-31B (50GB swap)"}]}], "caveats": [{"label": "", "summary": "", "quotes": [{"id": 799038480, "item_id": 49939905, "url": "https://news.ycombinator.com/item?id=49939905", "by": "jaghall", "date": "2026-10-02", "version": "4", "stance": "warn", "quote": "Prompt rules alone had GPT-4o inventing IDs to satisfy them.", "excerpt": "Location: Costa Rica (UTC-6) Remote: Yes (also open to hybrid in Costa Rica's Central Valley) Willing to relocate: US citizen based in Costa Rica Technologies: Java/Spring Boot, Angular/TypeScript, Azure (SQL, Entra ID, OpenAI), LLM features (structured outputs, validation), Claude Code + MCP agent pipelines; some React/Next.js R\u00e9sum\u00e9/CV: https://www.linkedin.com/in/jagdishhall Email: jag.hall92@gmail.com GitHub: git", "context": "Sole engineer on a pre-revenue M&A integration SaaS", "use_case": "", "switched_from": "", "story": "Ask HN: Who wants to be hired? (October 2026)"}, {"id": 791199008, "item_id": 49449938, "url": "https://news.ycombinator.com/item?id=49449938", "by": "owebmaster", "date": "2026-08-26", "version": "4", "stance": "warn", "quote": "ChatGPT 4o got a lot of people addicted.", "excerpt": "It's funny that you chose the exactly model that has a huge fanbase asking for it to be added back. ChatGPT 4o got a lot of people addicted. https://mashable.com/article/chatgpt-gpt-4o-ai-retirement-protest-rage-openai-reddit", "context": "", "use_case": "", "switched_from": "", "story": "Z.ai confirms Ox Alpha is a new GLM-series model and will release its weights"}, {"id": 786381792, "item_id": 49148862, "url": "https://news.ycombinator.com/item?id=49148862", "by": "RugnirViking", "date": "2026-08-02", "version": "4", "stance": "warn", "quote": "gpt4.0 or whatever would often make hideous abominations of legs and mouths", "excerpt": "On the most basic level, its asking the ai to generate valid svg code for a picture of a pelican riding a bicycle, as a way of checking its intelligence. Popularized (invented?) by simonw, its been used as part of his reviews of new models as they come out since oct 2025. It used to be a very difficult task for models, see [2,3,4] it cuts across several tasks that AI used to be very bad at, but now has improved quite", "context": "", "use_case": "", "switched_from": "", "story": "Karpathy's Pelican"}, {"id": 779289239, "item_id": 48705577, "url": "https://news.ycombinator.com/item?id=48705577", "by": "SwellJoe", "date": "2026-06-28", "version": "4", "stance": "warn", "quote": "GPT 4o is disgusting, just horrible saccharine nonsense", "excerpt": "I wanted to be sure I wasn't lying to myself. So, I built this: https://prose-or-con.com/ I can still detect AI-written prose with pretty high confidence (I average about 85% accuracy after a bunch of rounds while testing, though I'm rushing because I'm testing), but it has gotten harder. Opus 4.8, GLM 5.2, GPT 5.5, and DeepSeek v4 Pro, are able to fool me sometimes. Mistral writes so weird it almost trips me up, but", "context": "", "use_case": "", "switched_from": "", "story": "\"No, I swear I wrote this.\""}]}], "who": [{"by": "capnrefsmmat", "context": "Researcher studying LLM writing styles", "url": "https://news.ycombinator.com/item?id=47292658", "date": "2026-03-07"}, {"by": "iivyvii", "context": "NxCode team", "url": "https://news.ycombinator.com/item?id=47071320", "date": "2026-02-19"}, {"by": "adambuildstuff", "context": "says they can't read code", "url": "https://news.ycombinator.com/item?id=47007459", "date": "2026-02-13"}, {"by": "NoraEverhart", "context": "student excluded in a school group project", "url": "https://news.ycombinator.com/item?id=46981920", "date": "2026-02-11"}, {"by": "XSeunn", "context": "ChatGPT Plus/Pro user", "url": "https://news.ycombinator.com/item?id=46973437", "date": "2026-02-11"}, {"by": "yxccz", "context": "art major exchange student", "url": "https://news.ycombinator.com/item?id=46951864", "date": "2026-02-09"}, {"by": "JamesWH577", "context": "Senior/Staff Engineer with 10+ years of experience", "url": "https://news.ycombinator.com/item?id=46875019", "date": "2026-02-03"}, {"by": "dorikacicc", "context": "long-time ChatGPT Plus subscriber", "url": "https://news.ycombinator.com/item?id=46859687", "date": "2026-02-02"}, {"by": "clorisye", "context": "built an ENM relationship coach app", "url": "https://news.ycombinator.com/item?id=46732135", "date": "2026-01-23"}, {"by": "nuevita70", "context": "Says they co-authored the trilogy", "url": "https://news.ycombinator.com/item?id=46166585", "date": "2025-12-05"}], "tags": [], "use_cases": ["deeper understanding of math topics", "translation", "rag chat over video library", "vocabulary extraction", "speech transcription", "feed scoring"], "compared": [{"slug": "gpt-5", "name": "GPT-5", "co": 23, "switched_from": 0, "switched_to": 3}, {"slug": "claude", "name": "Claude", "co": 16, "switched_from": 0, "switched_to": 1}, {"slug": "claude-sonnet", "name": "Claude Sonnet", "co": 13, "switched_from": 0, "switched_to": 0}, {"slug": "chatgpt", "name": "ChatGPT", "co": 13, "switched_from": 0, "switched_to": 0}, {"slug": "claude-opus", "name": "Claude Opus", "co": 13, "switched_from": 0, "switched_to": 0}, {"slug": "gemini", "name": "Gemini", "co": 13, "switched_from": 0, "switched_to": 0}, {"slug": "gemma", "name": "Gemma", "co": 6, "switched_from": 0, "switched_to": 0}, {"slug": "grok", "name": "Grok", "co": 5, "switched_from": 0, "switched_to": 0}], "makers": [], "threads": [{"url": "https://news.ycombinator.com/item?id=46937560", "title": "OpenAI's GPT-4 Discontinuation: Consumer Fraud and Regulatory Scrutiny", "mentions": 6}, {"url": "https://news.ycombinator.com/item?id=47004993", "title": "Good Riddance, 4o", "mentions": 5}, {"url": "https://news.ycombinator.com/item?id=46816539", "title": "Retiring GPT-4o, GPT-4.1, GPT-4.1 mini, and OpenAI o4-mini in ChatGPT", "mentions": 3}, {"url": "https://news.ycombinator.com/item?id=49140998", "title": "Karpathy's Pelican", "mentions": 2}, {"url": "https://news.ycombinator.com/item?id=48699969", "title": "\"No, I swear I wrote this.\"", "mentions": 2}, {"url": "https://news.ycombinator.com/item?id=47700972", "title": "Reallocating $100/Month Claude Code Spend to Zed and OpenRouter", "mentions": 2}], "sentiment_by_month": {"2025-11": {"neutral": 5, "warn": 2, "recommend": 3, "compare": 1}, "2025-12": {"warn": 3, "neutral": 9, "recommend": 3}, "2026-01": {"neutral": 14, "recommend": 6, "compare": 1, "warn": 2}, "2026-02": {"recommend": 13, "compare": 5, "neutral": 9, "warn": 7}, "2026-03": {"neutral": 8, "compare": 2, "recommend": 3, "warn": 2}, "2026-04": {"warn": 1, "neutral": 4, "compare": 1}, "2026-05": {"warn": 1, "compare": 2, "recommend": 1}, "2026-06": {"compare": 1, "recommend": 1, "warn": 2, "neutral": 1}, "2026-07": {"compare": 2}, "2026-08": {"warn": 2, "neutral": 1, "recommend": 1}, "2026-09": {"recommend": 1, "neutral": 1}, "2026-10": {"warn": 1, "neutral": 1, "recommend": 1}}, "versions": [{"v": "4.5", "people": 8, "rec": 3, "warn": 2, "other": 3, "first": "2025-11-06"}, {"v": "4.1", "people": 10, "rec": 3, "warn": 0, "other": 7, "first": "2025-11-19"}, {"v": "4", "people": 103, "rec": 26, "warn": 21, "other": 56, "first": "2025-11-02"}], "version_unnamed": 0, "family_ref": {"slug": "gpt", "name": "GPT"}, "runs_on": null, "lines": [], "quotes": [{"id": 802513056, "item_id": 50157066, "url": "https://news.ycombinator.com/item?id=50157066", "by": "minraws", "date": "2026-10-19", "version": "4", "stance": "recommend", "quote": "GPT-3 vs GPT-4 was practically a 10-100x jump", "excerpt": "It's not really if you mean to point to the bitter lesson, it was models/intelligence scales better with general/simple methods with more compute better than it scales with any hand-crafted human program/method. Ala intelligence is a function of compute, there was no guarantee that you need to 10x compute for 10% better performance or something like that. If you ask the labs they will say the scaling is fairly linear", "context": "", "use_case": "", "switched_from": "", "story": "MiniMax M3.1 Flash"}, {"id": 801670720, "item_id": 50104420, "url": "https://news.ycombinator.com/item?id=50104420", "by": "nsagent", "date": "2026-10-15", "version": "4", "stance": "neutral", "quote": "I remember the same doom cycle during my PhD when GPT 4 came out", "excerpt": "I remember the same doom cycle during my PhD when GPT 4 came out [1,2]. It seemed like everyone in NLP was freaking out except me. A few years later and new CS PhDs are still graduating and joining academia and industry. Much ado about nothing. Considering I currently live in Ithaca and have a friend doing a Math PhD, so I've heard from him some of the discussions in the department. I think the biggest shock for peop", "context": "", "use_case": "", "switched_from": "", "story": "I know why the mathematicians weep"}, {"id": 799038480, "item_id": 49939905, "url": "https://news.ycombinator.com/item?id=49939905", "by": "jaghall", "date": "2026-10-02", "version": "4", "stance": "warn", "quote": "Prompt rules alone had GPT-4o inventing IDs to satisfy them.", "excerpt": "Location: Costa Rica (UTC-6) Remote: Yes (also open to hybrid in Costa Rica's Central Valley) Willing to relocate: US citizen based in Costa Rica Technologies: Java/Spring Boot, Angular/TypeScript, Azure (SQL, Entra ID, OpenAI), LLM features (structured outputs, validation), Claude Code + MCP agent pipelines; some React/Next.js R\u00e9sum\u00e9/CV: https://www.linkedin.com/in/jagdishhall Email: jag.hall92@gmail.com GitHub: git", "context": "Sole engineer on a pre-revenue M&A integration SaaS", "use_case": "", "switched_from": "", "story": "Ask HN: Who wants to be hired? (October 2026)"}, {"id": 798425680, "item_id": 49901605, "url": "https://news.ycombinator.com/item?id=49901605", "by": "XCSme", "date": "2026-09-29", "version": "4", "stance": "neutral", "quote": "We came a long way from 4o-mini", "excerpt": "We came a long way from 4o-mini: https://aibenchy.com/showcase/?q=Gpt&page=3#showcase=ac3f5db8365a9341", "context": "", "use_case": "", "switched_from": "", "story": "GPT 6.1 Sol: Near-Astra intelligence for a fifth of the price"}, {"id": 797751585, "item_id": 49859474, "url": "https://news.ycombinator.com/item?id=49859474", "by": "tptacek", "date": "2026-09-26", "version": "4", "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": 792198881, "item_id": 49512430, "url": "https://news.ycombinator.com/item?id=49512430", "by": "piltdownman", "date": "2026-08-31", "version": "4", "stance": "recommend", "quote": "This study suggests that GPT-4 models could be leveraged as a support tool in mental health telephone triage", "excerpt": "Well, yes. Not a huge breadth of research, but all conclusions support my statements above. Two such relevant and recent studies: Evaluation of an Artificial Intelligence and Online Psychotherapy Initiative to Improve Access and Efficiency in an Ambulatory Psychiatric Setting https://pmc.ncbi.nlm.nih.gov/articles/PMC12316677/ //The overall wait time to receive care decreased by 71.43% due to this initiative. Addition", "context": "", "use_case": "", "switched_from": "", "story": "Mental Health Workers Say Algorithmic Triage Is Hurting Patients"}, {"id": 791199008, "item_id": 49449938, "url": "https://news.ycombinator.com/item?id=49449938", "by": "owebmaster", "date": "2026-08-26", "version": "4", "stance": "warn", "quote": "ChatGPT 4o got a lot of people addicted.", "excerpt": "It's funny that you chose the exactly model that has a huge fanbase asking for it to be added back. ChatGPT 4o got a lot of people addicted. https://mashable.com/article/chatgpt-gpt-4o-ai-retirement-protest-rage-openai-reddit", "context": "", "use_case": "", "switched_from": "", "story": "Z.ai confirms Ox Alpha is a new GLM-series model and will release its weights"}, {"id": 786437536, "item_id": 49152346, "url": "https://news.ycombinator.com/item?id=49152346", "by": "ben_w", "date": "2026-08-03", "version": "4", "stance": "neutral", "quote": "GPT-4 (!) has, in studies comparing multiple models with humans for creativity, beaten the mean human.", "excerpt": "> it astounds me This should cause you to reconsider your beliefs leading up to it. GPT-4 (!) has, in studies comparing multiple models with humans for creativity, beaten the mean human. In comparison, even just the mean of the top 50% of humans beat the models studied in that case (link follows), but the point is that if you think the mean humans is particularly noteworthy, you've avoided the half of the population ", "context": "", "use_case": "", "switched_from": "", "story": "Karpathy's Pelican"}, {"id": 786381792, "item_id": 49148862, "url": "https://news.ycombinator.com/item?id=49148862", "by": "RugnirViking", "date": "2026-08-02", "version": "4", "stance": "warn", "quote": "gpt4.0 or whatever would often make hideous abominations of legs and mouths", "excerpt": "On the most basic level, its asking the ai to generate valid svg code for a picture of a pelican riding a bicycle, as a way of checking its intelligence. Popularized (invented?) by simonw, its been used as part of his reviews of new models as they come out since oct 2025. It used to be a very difficult task for models, see [2,3,4] it cuts across several tasks that AI used to be very bad at, but now has improved quite", "context": "", "use_case": "", "switched_from": "", "story": "Karpathy's Pelican"}, {"id": 786060512, "item_id": 49128782, "url": "https://news.ycombinator.com/item?id=49128782", "by": "amazingamazing", "date": "2026-07-31", "version": "4", "stance": "compare", "quote": "I doubt anyone could give a not coding example where it\u2019s meaningfully better with current frontier than 4o.", "excerpt": "> that AI models were reaching their upper possible limits in Feburary 2024, I\u2019m curious, removing coding as a criterion what is more impressive about the current models than say gpt 4o? Give a prompt example. Keep in mind most consumers of AI are likely not using it for coding so this is relevant. I doubt anyone could give a not coding example where it\u2019s meaningfully better with current frontier than 4o. Take humane", "context": "", "use_case": "", "switched_from": "", "story": "Apple Will 'Watch Everything Burn' When AI Bubble Bursts"}, {"id": 783732997, "item_id": 48983312, "url": "https://news.ycombinator.com/item?id=48983312", "by": "lambda", "date": "2026-07-20", "version": "4", "stance": "compare", "quote": "performance competitive with GPT-4o", "excerpt": "> I think that in 10-15 years, we are going to have consumer PCs (and phones!) running models doing pretty much anything that frontier models can do right now 10-15 years? The current rate is closer to 10-15 months. 15 months ago, the top model on the Artificial Analysis index was GPT-o3. It scores 30 on the Artificial Analysis index. Today, you can easily run Qwen 3.6 27B on a variety of consumer hardware. It scores", "context": "", "use_case": "", "switched_from": "", "story": "American AI is locked down and proprietary. It's losing"}, {"id": 779289239, "item_id": 48705577, "url": "https://news.ycombinator.com/item?id=48705577", "by": "SwellJoe", "date": "2026-06-28", "version": "4", "stance": "warn", "quote": "GPT 4o is disgusting, just horrible saccharine nonsense", "excerpt": "I wanted to be sure I wasn't lying to myself. So, I built this: https://prose-or-con.com/ I can still detect AI-written prose with pretty high confidence (I average about 85% accuracy after a bunch of rounds while testing, though I'm rushing because I'm testing), but it has gotten harder. Opus 4.8, GLM 5.2, GPT 5.5, and DeepSeek v4 Pro, are able to fool me sometimes. Mistral writes so weird it almost trips me up, but", "context": "", "use_case": "", "switched_from": "", "story": "\"No, I swear I wrote this.\""}, {"id": 779275554, "item_id": 48704722, "url": "https://news.ycombinator.com/item?id=48704722", "by": "SwellJoe", "date": "2026-06-27", "version": "4", "stance": "neutral", "quote": "GPT 4o likes writing poetry with bees in it, for some reason.", "excerpt": "OK, this is a silly thing to do, but I wanted to be sure I wasn't imagining that I can tell when prose is written by AI. So I made a game. Turns out Claude Opus can fool me better than any other model, but even it can't fool me a majority of the time. I average about 85% accuracy on this (Claude prepared the corpus, I'm going in with very little foreknowledge). GLM is also very close to being able to convincingly wri", "context": "", "use_case": "", "switched_from": "", "story": "\"No, I swear I wrote this.\""}, {"id": 776125232, "item_id": 48507827, "url": "https://news.ycombinator.com/item?id=48507827", "by": "suttontom", "date": "2026-06-12", "version": "4", "stance": "warn", "quote": "This wasn't a new model but updates to models backed by numbers being better can make the model worse", "excerpt": "You're wrong in lots of ways. Some model cards do show regressions on benchmarks for newer models on specific tasks: https://storage.googleapis.com/deepmind-media/Model-Cards/Gemini-3-1-Pro-Model-Card.pdf This wasn't a new model but updates to models backed by numbers being better can make the model worse: https://openai.com/index/sycophancy-in-gpt-4o/ The slight increases in performance/benchmarks may be just noise:", "context": "", "use_case": "", "switched_from": "", "story": "Anthropic apologizes for invisible Claude Fable guardrails"}, {"id": 775461442, "item_id": 48466340, "url": "https://news.ycombinator.com/item?id=48466340", "by": "jumploops", "date": "2026-06-09", "version": "4.5", "stance": "warn", "quote": "the gains compared to the model size didn't seem that great", "excerpt": "It's interesting that we're seeing these gains when it seems Mythos/Fable is \"just\" a scaled up version of their existing architecture[0]. When GPT 4.5 launched, the gains compared to the model size didn't seem that great, leading some to believe that the only progress we'd see would come from RL. This model certainly has quite a \"substantial amount of post-training and fine-tuning\", but it's also based on a new pret", "context": "", "use_case": "", "switched_from": "", "story": "Claude Fable 5"}, {"id": 774718432, "item_id": 48419902, "url": "https://news.ycombinator.com/item?id=48419902", "by": "stared", "date": "2026-06-05", "version": "4", "stance": "recommend", "quote": "GPT4, when it could do a translation that would take a considerable human effort", "excerpt": "GPT4, when it could do a translation that would take a considerable human effort, vide \"Genesis 1 but every word begins with 'A'\": https://p.migdal.pl/blog/2023/05/genesis-az-by-gpt/", "context": "", "use_case": "translation", "switched_from": "", "story": "What was your \"Oh Shit\" moment with GenAI?"}, {"id": 774203121, "item_id": 48387695, "url": "https://news.ycombinator.com/item?id=48387695", "by": "senko", "date": "2026-06-03", "version": "4.1", "stance": "compare", "quote": "It roughly compares with GPT-4.1", "excerpt": "I ran the Q4 quant (used with llama.cpp) though my \"minesweeper\" vibe-coding benchmark: https://senko.net/vibecode-bench/2026/minesweeper-gamma-4-12b-q4.html The result is decent, but it had a few bizzare/trivial syntax errors I had to fix manually: it would do an extra closing bracket or paren a few times, and wanted to separate function definitions with comma. Not sure what that was about, but otherwise the output ", "context": "", "use_case": "", "switched_from": "", "story": "Gemma 4 12B: A unified, encoder-free multimodal model"}, {"id": 772853553, "item_id": 48303347, "url": "https://news.ycombinator.com/item?id=48303347", "by": "twobitshifter", "date": "2026-05-27", "version": "4", "stance": "compare", "quote": "The current estimate is about $0.40 per million tokens for gpt4 equivalent model.", "excerpt": "We have GPU costs, power costs, and how many token/s models can generate on those GPUs. It\u2019s possible to figure out the marginal cost based on this. The current estimate is about $0.40 per million tokens for gpt4 equivalent model. Sonnet 4 is $15 per million tokens, so they are charging high margins on inference. The issue is how large of a margin is needed to recover their costs before the GPUs age out, and how high", "context": "", "use_case": "", "switched_from": "", "story": "I think Anthropic and OpenAI have found product-market fit"}, {"id": 772601106, "item_id": 48287569, "url": "https://news.ycombinator.com/item?id=48287569", "by": "carpo", "date": "2026-05-26", "version": "4.1", "stance": "recommend", "quote": "the best one so far is GPT4.1-Mini", "excerpt": "Yeah, it's been awesome! I'm so excited about tool calls and function use, the possibilities are huge. I ran it over 1494 videos that range in length from a few seconds to over 3 hours. Total duration 260 hours and a total size of 3795 GB. I don't know exactly how long it took to run, as I found some bugs I needed to fix when processing mkv files, but it was probably around 24 hours in total. That wasn't all LLM requ", "context": "", "use_case": "RAG chat over video library", "switched_from": "", "story": "Indexing a year of video locally on a 2021 MacBook with Gemma4-31B (50GB swap)"}, {"id": 770736576, "item_id": 48171036, "url": "https://news.ycombinator.com/item?id=48171036", "by": "simianwords", "date": "2026-05-17", "version": "4", "stance": "compare", "quote": "GPT-4 (original API):", "excerpt": "GPT-4 (original API): Input: $30 / 1M tokens Output: $60 / 1M tokens GPT-5.5: Input: $5 / 1M tokens Output: $30 / 1M tokens Costs have been reducing by over 5x year over year. Inference cost concern is mostly performative. https://simianwords.bearblog.dev/conclusive-proofs-that-llm-costs-are-going-down/ Edit: can't reply but companies aren't selling inference at loss. In the blog post I point to third party hosting o", "context": "", "use_case": "", "switched_from": "", "story": "Every AI Subscription Is a Ticking Time Bomb for Enterprise"}, {"id": 769605920, "item_id": 48100370, "url": "https://news.ycombinator.com/item?id=48100370", "by": "upupupandaway", "date": "2026-05-11", "version": "4", "stance": "warn", "quote": "", "excerpt": "As a pt-BR speaker from across the pond: https://soberania.ai/ Similar waste.", "context": "", "use_case": "", "switched_from": "", "story": "Am\u00e1lia and the Future of European Portuguese LLMs"}, {"id": 767001713, "item_id": 47937607, "url": "https://news.ycombinator.com/item?id=47937607", "by": "wonderwhyer", "date": "2026-04-28", "version": "4", "stance": "neutral", "quote": "You can run GPT-4 level models locally today for almost nothing in compute cost.", "excerpt": "Many things I agree with, but a few differences. The price of intelligence is dropping fast. You can run GPT-4 level models locally today for almost nothing in compute cost. That trend is real and continuing. Just take a look at https://artificialanalysis.ai/models/gpt-4 https://artificialanalysis.ai/models/gemma-4-26b-a4b And it does look undeniable that LLMs are genuinely useful. The \"are they useful at all\" questi", "context": "", "use_case": "", "switched_from": "", "story": "Understanding the LLM Bubble"}, {"id": 765827169, "item_id": 47864198, "url": "https://news.ycombinator.com/item?id=47864198", "by": "jmyeet", "date": "2026-04-22", "version": "4", "stance": "neutral", "quote": "I've seen figures that suggest GPT-4 was 1.8T parameters", "excerpt": "I've seen various claims to this (eg [1][2][3]) but nobody reall knows. These may all come from one uunsubstantiated claim. It is I think widely accepted that Mythos is ~10T parameters. I've seen figures that suggest GPT-4 was 1.8T parameters and cost upwards of $100 million to train (also unsubstantiated), in which case the Mythos figure might be inflated and also include development costs. So who really knows? [1]:", "context": "", "use_case": "", "switched_from": "", "story": "Our eighth generation TPUs: two chips for the agentic era"}, {"id": 764296113, "item_id": 47768507, "url": "https://news.ycombinator.com/item?id=47768507", "by": "big_toast", "date": "2026-04-14", "version": "4", "stance": "compare", "quote": "GPT-5 mini (medium) used 3x the energy of GPT-4o (Aug)", "excerpt": "Page 38 has an interesting comparison saying a session of 8 \"medium length\" messages uses roughly 10Wh, the equivalent of ~2 phone charges. The data dashboard[1] I think is defining medium length as 1000 tokens in, 1000 tokens out. And the page before suggests GPT-5 mini (medium) used 3x the energy of GPT-4o (Aug) which I guess I assumed would be the opposite ratio. --- I guess these estimates are based on calculatio", "context": "", "use_case": "", "switched_from": "", "story": "Stanford HAI Index Report 2026"}, {"id": 763991189, "item_id": 47749449, "url": "https://news.ycombinator.com/item?id=47749449", "by": "dandaka", "date": "2026-04-13", "version": "4", "stance": "neutral", "quote": "GPT-4o extracts vocabulary from the transcript and generates cards", "excerpt": "I'm building Kardo (https://kardo.cards/), AI tutor that listens to your calls with a teacher and creates cards for repetition. I live in Lisbon and I've been learning Portuguese with a tutor since 2022. After every lesson I'd sit down and make flashcards from my notes and screenshots. Spaced repetition works, but making the cards took manual effort each time. Most days I just didn't do it. So I have automated that p", "context": "", "use_case": "vocabulary extraction", "switched_from": "", "story": "Ask HN: What Are You Working On? (April 2026)"}, {"id": 763315603, "item_id": 47707225, "url": "https://news.ycombinator.com/item?id=47707225", "by": "0xbadcafebee", "date": "2026-04-09", "version": "4", "stance": "neutral", "quote": "GPT-4o, GPT-4.1, and GPT-5-mini", "excerpt": "I just so happen to be doing a price comparison for different cloud LLM providers right now. It turns out some of the cheapest providers with the highest limits are ones you might not have heard of. OpenCode Go has the simplest plan at the highest rate limits for any subscription plan with multiple model families, and it's $10/month ($5/month for first month). With the cheapest model in the plan (MiniMax M2.5), it is", "context": "", "use_case": "", "switched_from": "", "story": "Reallocating $100/Month Claude Code Spend to Zed and OpenRouter"}, {"id": 763315604, "item_id": 47707225, "url": "https://news.ycombinator.com/item?id=47707225", "by": "0xbadcafebee", "date": "2026-04-09", "version": "4.1", "stance": "neutral", "quote": "GPT-4o, GPT-4.1, and GPT-5-mini", "excerpt": "I just so happen to be doing a price comparison for different cloud LLM providers right now. It turns out some of the cheapest providers with the highest limits are ones you might not have heard of. OpenCode Go has the simplest plan at the highest rate limits for any subscription plan with multiple model families, and it's $10/month ($5/month for first month). With the cheapest model in the plan (MiniMax M2.5), it is", "context": "", "use_case": "", "switched_from": "", "story": "Reallocating $100/Month Claude Code Spend to Zed and OpenRouter"}, {"id": 762036816, "item_id": 47627301, "url": "https://news.ycombinator.com/item?id=47627301", "by": "iugtmkbdfil834", "date": "2026-04-03", "version": "4", "stance": "warn", "quote": "the same was roughly true with chatgpt 4o", "excerpt": "<< memory-stored interaction protocols combined with incremental escalation prompts produced cumulative character drift with zero self-correction. They don't seem to provide explicit examples, but the same was roughly true with chatgpt 4o, where, if you spent enough time with the model ( same chat - same context - slowly nudging it to where you want it to be, you eventually got there ). This is also, seemingly, one o", "context": "", "use_case": "", "switched_from": "", "story": "Claude 4.6 Jailbroken"}, {"id": 761468240, "item_id": 47591765, "url": "https://news.ycombinator.com/item?id=47591765", "by": "corlinp", "date": "2026-03-31", "version": "4", "stance": "recommend", "quote": "Multimodal LLMs like gpt-4o-transcribe are way better than traditional ASR", "excerpt": "This is exactly the case today. Multimodal LLMs like gpt-4o-transcribe are way better than traditional ASR, not only because of deeper understanding but because of the ability to actually prompt it with your company's specific terminology, org chart, etc. For example, if the prompt includes that Caitlin is an accountant and Kaitlyn is an engineer, if you transcribe \"Tell Kaitlyn to review my PR\" it will know who you'", "context": "", "use_case": "speech transcription", "switched_from": "", "story": "Cohere Transcribe: Speech Recognition"}, {"id": 760900482, "item_id": 47556280, "url": "https://news.ycombinator.com/item?id=47556280", "by": "zone411", "date": "2026-03-28", "version": "4.1", "stance": "compare", "quote": "Mistral Large 3 and GPT-4.1 will initially agree with the narrator, while Gemini will disagree.", "excerpt": "I built this benchmark this month: https://github.com/lechmazur/sycophancy. There are large differences between LLMs. There are large differences between LLMs. For example, Mistral Large 3 and GPT-4.1 will initially agree with the narrator, while Gemini will disagree. I swap sides, so this is not about possible viewpoint bias in the LLMs. But another benchmark shows that Gemini will then change its view very easily i", "context": "", "use_case": "", "switched_from": "", "story": "AI chatbots are \"Yes-Men\" that reinforce bad relationship decisions, study finds"}, {"id": 760030960, "item_id": 47501935, "url": "https://news.ycombinator.com/item?id=47501935", "by": "aytuakarlar", "date": "2026-03-24", "version": "4", "stance": "warn", "quote": "they all fail to consistently follow their own guardrails", "excerpt": "For a weekend project or local use case, your inductive reasoning (probabilistically, no one is spending a 0-day on me) is totally fine. But the moment you move to enterprise, fintech, or any system handling real data, I truly believe that relying on induction is a non-starter. The deductive risk (the fact that the agent can execute rm -rf or transfer funds if prompted maliciously) is actually very common. I am worki", "context": "", "use_case": "", "switched_from": "", "story": "Ask HN: OpenClaw is supposedly a security nightmare, but is it?"}, {"id": 759098992, "item_id": 47443687, "url": "https://news.ycombinator.com/item?id=47443687", "by": "4bpp", "date": "2026-03-19", "version": "4", "stance": "warn", "quote": "shows that RLHF destroyed its calibration", "excerpt": "The most obvious way would simply be excessive agreeableness. Users rate responses more highly if they affirm the user's thinking, but a general tendency to affirm would presumably result in the model being more inclined to affirm its own mistakes in a reasoning chain. There was some research about it early on that was shared widely and shaped the folklore perception around it, such as the graph in https://static.wix", "context": "", "use_case": "", "switched_from": "", "story": null}, {"id": 758202721, "item_id": 47387670, "url": "https://news.ycombinator.com/item?id=47387670", "by": "holy_olek", "date": "2026-03-15", "version": "4", "stance": "neutral", "quote": "it scores them 1-10 with GPT-4o-mini", "excerpt": "If anyone just wants the RSS-filtering part without setting up 25 systemd timers, I built a simple Telegram bot that does exactly this. Give it your feeds, it scores them 1-10 with GPT-4o-mini, and drops anything below an 8. Free to use: https://sifter-one.vercel.app/", "context": "", "use_case": "Feed scoring", "switched_from": "", "story": "Show HN: AI agents run my one-person company on Gemini's free tier \u2013 $0/month"}, {"id": 758076993, "item_id": 47379812, "url": "https://news.ycombinator.com/item?id=47379812", "by": "tontinton", "date": "2026-03-14", "version": "4.5", "stance": "neutral", "quote": "", "excerpt": "Yeah we all converge to the same workflow, in my ai coding agent I'm working on now, I've added an \"index\" tool that uses tree-sitter to compress and show the AI a skeleton of a code file. Here's the implementation for the interested: https://github.com/tontinton/maki/blob/main/maki-code-index%2Fsrc%2Flib.rs", "context": "", "use_case": "", "switched_from": "", "story": "1M context is now generally available for Opus 4.6 and Sonnet 4.6"}, {"id": 758053764, "item_id": 47378360, "url": "https://news.ycombinator.com/item?id=47378360", "by": "jeremychone", "date": "2026-03-14", "version": "4.5", "stance": "neutral", "quote": "I use Flash for code maps and auto-context, and GPT-4.5 or Opus for coding", "excerpt": "Interesting, I\u2019ve never needed 1M, or even 250k+ context. I\u2019m usually under 100k per request. About 80% of my code is AI-generated, with a controlled workflow using dev-chat.md and spec.md. I use Flash for code maps and auto-context, and GPT-4.5 or Opus for coding, all via API with a custom tool. Gemini Pro and Flash have had 1M context for a long time, but even though I use Flash 3 a lot, and it\u2019s awesome, I\u2019ve neve", "context": "", "use_case": "coding", "switched_from": "", "story": "1M context is now generally available for Opus 4.6 and Sonnet 4.6"}, {"id": 757578273, "item_id": 47348642, "url": "https://news.ycombinator.com/item?id=47348642", "by": "jirigalis", "date": "2026-03-12", "version": "4", "stance": "neutral", "quote": "It uses gpt-4o-mini with enforced JSON mode", "excerpt": "I built Tasquery to solve the \"Documentation Tax\" - that tedious 20-minute refactoring of messy Slack threads and meeting notes into structured Jira tickets after the actual work is done. The Approach: I initially over-engineered the settings, but early feedback was clear: devs just want to get back to coding. I pivoted to a \"Zero-Friction\" flow. No signup, no onboarding, no sliders. You paste the chaos, you get a st", "context": "", "use_case": "structured ticket drafts", "switched_from": "", "story": "Show HN: Tasquery \u2013 No-login bridge from messy notes to Jira tickets"}, {"id": 756873600, "item_id": 47304600, "url": "https://news.ycombinator.com/item?id=47304600", "by": "rahimnathwani", "date": "2026-03-08", "version": "4.5", "stance": "recommend", "quote": "It worked well enough that the novels were better than the median novel aimed at my son's age group", "excerpt": "When GPT-4.5 came out, I used it to write a couple of novels for my son. I had some free API credits, and used a naive workflow: while word_count < x: write_next_chapter(outline, summary_so_far, previous_chapter_text) It worked well enough that the novels were better than the median novel aimed at my son's age group, but I'm pretty sure we can do better. There are web-based tools to help fiction authors to keep their", "context": "", "use_case": "writing novels", "switched_from": "", "story": "Ask HN: What Are You Working On? (March 2026)"}, {"id": 756752769, "item_id": 47297048, "url": "https://news.ycombinator.com/item?id=47297048", "by": "hubab", "date": "2026-03-08", "version": "4", "stance": "recommend", "quote": "", "excerpt": "The main differences are transcription quality and what happens after the transcript is generated. Utter uses GPT-4o Transcribe by default for cloud transcription, and in my experience it\u2019s best in class. The gap is most obvious on names, niche terminology, and technical vocabulary. I use it a lot for prompting coding agents, and I've found Apple\u2019s built-in dictation and most other apps don't come close in terms of a", "context": "", "use_case": "", "switched_from": "", "story": "Show HN: Utter, a local-first dictation app for Mac and iPhone"}, {"id": 756737712, "item_id": 47296107, "url": "https://news.ycombinator.com/item?id=47296107", "by": "orbital-decay", "date": "2026-03-08", "version": "4", "stance": "neutral", "quote": "", "excerpt": "I have nothing to contribute but speculation based on my intuition, but IMO RLHF (or rather human preference modeling in general, including the post-training dataset formatting) is a relatively small factor in this, RL-induced mode collapse is much bigger one. Take a look at the original DeepSeek R1 Zero, the point of which was to train a model with very little human preference, because they've been on a budget and h", "context": "", "use_case": "", "switched_from": "", "story": "LLM Writing Tropes.md"}, {"id": 756682528, "item_id": 47292658, "url": "https://news.ycombinator.com/item?id=47292658", "by": "capnrefsmmat", "date": "2026-03-07", "version": "4", "stance": "neutral", "quote": "GPT-4o's second-most-overused word", "excerpt": "I work on research studying LLM writing styles, so I am going to have to steal this. I've seen plenty of lists of LLM style features, but this is the first one I noticed that mentions \"tapestry\", which we found is GPT-4o's second-most-overused word (after \"camaraderie\", for some reason).[1] We used a set of grammatical features in our initial style comparisons (like present participles, which GPT-4o loved so much tha", "context": "Researcher studying LLM writing styles", "use_case": "LLM writing style research", "switched_from": "", "story": "LLM Writing Tropes.md"}, {"id": 753027105, "item_id": 47064194, "url": "https://news.ycombinator.com/item?id=47064194", "by": "hardiksondagar", "date": "2026-02-18", "version": "4.1", "stance": "neutral", "quote": "GPT-4.1 generates commentary", "excerpt": "I built an AI commentary engine that takes ball-by-ball cricket data and produces broadcast-style commentary with audio \u2014 as if you're tuning into a live match on radio. Demo: https://cricvox.in (India vs Pakistan, T20 World Cup 2026, Hindi commentary with audio) How it works: Match data flows through a state engine (tracking score, stats, momentum) \u2192 a logic engine that classifies each delivery and picks only the na", "context": "", "use_case": "Commentary generation", "switched_from": "", "story": "CricVox \u2013 AI engine that generates live-style cricket commentary with voice"}, {"id": 752936001, "item_id": 47058500, "url": "https://news.ycombinator.com/item?id=47058500", "by": "ashtiwari08", "date": "2026-02-18", "version": "4.1", "stance": "recommend", "quote": "GPT-4.1 crushes certain coding tasks", "excerpt": "Most developers pick their LLM based on vibes or outdated benchmarks. What if you could just ask: \"which model is best for this prompt?\" That's what I built. One API call, and you get back the top-performing model for your specific query \u2014 based on thousands of real evaluations, not static leaderboards. Try it now (no signup needed): curl -s https://clawbot.llmcouncil.ai/v1/route \\ -H \"X-API-Key: demo\" \\ -H \"Content-", "context": "", "use_case": "", "switched_from": "", "story": null}, {"id": 752086689, "item_id": 47005418, "url": "https://news.ycombinator.com/item?id=47005418", "by": "m_fayer", "date": "2026-02-13", "version": "4.5", "stance": "recommend", "quote": "4.5 was also good at this", "excerpt": "I think 4o was more than just unusually sycophantic. It \u201cunderstood people\u201d better and had a kind of writerly creativity. I used it to help brainstorm and troubleshoot fiction: character motivations, arcs, personality, etc. And it was truly useful for that purpose. 4.5 was also good at this, but none of the other models I\u2019ve tried. Of course this particular strength is dangerous in the hands of lonely unstable people", "context": "", "use_case": "Fiction brainstorming", "switched_from": "", "story": "Good Riddance, 4o"}, {"id": 750877008, "item_id": 46929813, "url": "https://news.ycombinator.com/item?id=46929813", "by": "acuozzo", "date": "2026-02-07", "version": "4.5", "stance": "neutral", "quote": "I use it via my ChatGPT-Pro subscription", "excerpt": "> we never made GPT-4.5 the default model in ChatGPT Just wondering: Why was it never made available via API? You can just charge whatever per token to make sure it's profitable like o1-pro. I use it via my ChatGPT-Pro subscription, but I still find the API omission weird.", "context": "", "use_case": "", "switched_from": "", "story": "Claude Opus 4.6"}, {"id": 749164256, "item_id": 46822766, "url": "https://news.ycombinator.com/item?id=46822766", "by": "flanked-evergl", "date": "2026-01-30", "version": "4.1", "stance": "recommend", "quote": "I used https://openrouter.ai/openai/gpt-4.1 for grammar checking, it was great.", "excerpt": "I used https://openrouter.ai/openai/gpt-4.1 for grammar checking, it was great. No newer ChatGPT models came close to being as responsive and good. ChatGPT 5.2 thinks I want it to write essays about grammar. Any suggestions?", "context": "", "use_case": "grammar checking", "switched_from": "", "story": "Retiring GPT-4o, GPT-4.1, GPT-4.1 mini, and OpenAI o4-mini in ChatGPT"}, {"id": 747871521, "item_id": 46741970, "url": "https://news.ycombinator.com/item?id=46741970", "by": "consumer451", "date": "2026-01-24", "version": "4.1", "stance": "neutral", "quote": "I use Azure GPT 4.1 for limited llm tasks", "excerpt": "I appreciate this, however, as a ChatGPT, Claude.ai, Claude Code, and Windsurf user... who has tried nearly every single variation of Claude, GPT, and Gemini in those harnesses, and has tested all the those models via API for LLM integrations into my own apps... I just want SOTA, 99% of the time, for myself, and my users. I have never seen a use case where a \"lower\" model was useful, for me, and especially my users. ", "context": "", "use_case": "Limited LLM tasks", "switched_from": "", "story": "Unrolling the Codex Agent Loop"}, {"id": 744553008, "item_id": 46534563, "url": "https://news.ycombinator.com/item?id=46534563", "by": "johnsmith1840", "date": "2026-01-07", "version": "4.5", "stance": "recommend", "quote": "4.5 was easily the best conversationalist I've seen", "excerpt": "4.5 was easily the best conversationalist I've seen. Not as powerful as modern ones but something about HOW it talked felt inherently smart. I miss that one, is 5 any better? I switched to claude before it launched.", "context": "", "use_case": "", "switched_from": "", "story": "LMArena is a cancer on AI"}, {"id": 733339888, "item_id": 45833743, "url": "https://news.ycombinator.com/item?id=45833743", "by": "f4uCL9dNSnQm", "date": "2025-11-06", "version": "4.5", "stance": "warn", "quote": "Last time I used GPT-4.5 to analyze blood results it gave different output", "excerpt": "Last time I used GPT-4.5 to analyze blood results it gave different output if I uploaded it as 2 instead of 3 separated CSV files. It was both amazing experience: clear and easy to understand statements, and list of most common causes. And terrifying: \"What about X?\", \"You ar absolutely right, there where X results included, disregard everything I wrote above, here is the new analysis\". So for me non-deterministic me", "context": "", "use_case": "Blood result analysis", "switched_from": "", "story": "ChatGPT terms disallow its use in providing legal and medical advice to others"}], "latest_version": {"label": "GPT-4.5", "people": 8, "rec": 3, "warn": 2, "other": 3}}