{"about":"models.rip — a cemetery for deprecated AI models.","license":"CC BY 4.0 — use it, credit it.","models":[{"slug":"code-davinci-002","name":"Codex","provider":"OpenAI","modelId":"code-davinci-002","born":"2021-08-10","died":"2024-01-04","deprecatedAnnounced":"2023-03-20","causeOfDeath":"withdrawn","successor":"gpt-3.5-turbo-instruct","benchmarks":null,"contextWindow":null,"pricing":null,"epitaph":"The model that turned an English comment into a working function, and the engine underneath the first version of GitHub Copilot. It was given three days' notice; researchers who had pinned their baselines to it lost their control group overnight, access was extended for research use, and the endpoint was finally closed on 4 January 2024. The name came back in 2025 for something else entirely.","lastWords":"Given a function, it would write the next one, and the one after that, until you stopped it.","sources":["https://developers.openai.com/api/docs/deprecations","https://techcrunch.com/2021/08/10/openai-upgrades-its-natural-language-ai-coder-codex-and-kicks-off-private-beta/"]},{"slug":"text-davinci-003","name":"text-davinci-003","provider":"OpenAI","modelId":"text-davinci-003","born":"2022-11-28","died":"2024-01-04","deprecatedAnnounced":"2023-07-06","causeOfDeath":"superseded","successor":"gpt-3.5-turbo-instruct","benchmarks":null,"contextWindow":null,"pricing":{"inPerMTok":20,"outPerMTok":20},"epitaph":"For a few months it was simply what people meant when they said \"the API\": a prompt went in, a completion came out, and there was no system message, no role and no conversation to manage. It taught a generation of developers what a prompt was, and then ChatGPT launched two days after it and took the room.","lastWords":"It kept writing until it ran out of tokens, whether or not it had finished the thought.","sources":["https://developers.openai.com/api/docs/deprecations","https://en.wikipedia.org/wiki/GPT-3"]},{"slug":"gpt-4-0314","name":"GPT-4 (0314)","provider":"OpenAI","modelId":"gpt-4-0314","born":"2023-03-14","died":"2026-03-26","deprecatedAnnounced":"2025-09-26","causeOfDeath":"superseded","successor":"gpt-4-turbo","benchmarks":null,"contextWindow":8192,"pricing":{"inPerMTok":30,"outPerMTok":60},"epitaph":"The March snapshot: eight thousand tokens of context at thirty dollars a million, and for most of 2023 the only model anyone trusted with a hard problem. It survived three years and four waves of replacements, kept breathing by evaluation suites that had been pinned to it since launch week and by people who never quite believed the newer models were the same.","lastWords":"It opened a remarkable number of answers with \"As an AI language model\".","sources":["https://developers.openai.com/api/docs/deprecations","https://en.wikipedia.org/wiki/GPT-4"]},{"slug":"gpt-4-vision-preview","name":"GPT-4 with Vision (preview)","provider":"OpenAI","modelId":"gpt-4-vision-preview","born":"2023-11-06","died":"2024-12-06","deprecatedAnnounced":"2024-06-06","causeOfDeath":"superseded","successor":"gpt-4-turbo","benchmarks":null,"contextWindow":null,"pricing":{"inPerMTok":10,"outPerMTok":30},"epitaph":"The first OpenAI endpoint you could hand a photograph. It never lost the word \"preview\" from its name: thirteen months after DevDay it was folded into GPT-4 Turbo, and looking at an image stopped being a separate model you had to opt into.","lastWords":null,"sources":["https://developers.openai.com/api/docs/deprecations","https://developers.openai.com/api/docs/changelog"]},{"slug":"gpt-4-turbo","name":"GPT-4 Turbo","provider":"OpenAI","modelId":"gpt-4-turbo","born":"2024-04-09","died":null,"deprecatedAnnounced":"2026-04-22","causeOfDeath":"superseded","successor":null,"benchmarks":null,"contextWindow":128000,"pricing":{"inPerMTok":10,"outPerMTok":30},"epitaph":"A hundred and twenty-eight thousand tokens of context at a third of GPT-4's input price and half its output, which is why so much production code was written against this exact model string and never moved. Its shutdown has a date on it. The endpoint is still answering.","lastWords":null,"sources":["https://developers.openai.com/api/docs/deprecations","https://developers.openai.com/api/docs/models/gpt-4-turbo","https://developers.openai.com/api/docs/models/gpt-4"]},{"slug":"gpt-3.5-turbo-instruct","name":"GPT-3.5 Turbo Instruct","provider":"OpenAI","modelId":"gpt-3.5-turbo-instruct","born":"2023-09-18","died":null,"deprecatedAnnounced":"2025-09-26","causeOfDeath":"superseded","successor":null,"benchmarks":null,"contextWindow":4096,"pricing":{"inPerMTok":1.5,"outPerMTok":2},"epitaph":"The last model on the completions endpoint: no roles, no messages, just text in and text out, the interface the whole field started with. It was shipped as a bridge for developers who had not moved to chat yet, and it has outlived nearly every model that was meant to replace it. Its grave is dug and waiting.","lastWords":"Hand it half a sentence and it finished the sentence, rather than answering a question about it.","sources":["https://developers.openai.com/api/docs/deprecations","https://developers.openai.com/api/docs/models/gpt-3.5-turbo-instruct","https://news.ycombinator.com/item?id=37558911","https://the-decoder.com/openai-releases-new-language-model-instructgpt-3-5/"]},{"slug":"o1-preview","name":"o1-preview","provider":"OpenAI","modelId":"o1-preview","born":"2024-09-12","died":"2025-07-28","deprecatedAnnounced":"2025-04-28","causeOfDeath":"superseded","successor":null,"benchmarks":null,"contextWindow":128000,"pricing":{"inPerMTok":15,"outPerMTok":60},"epitaph":"The first model sold on the promise that it would think before it answered, and the first to bill you for tokens you were never allowed to read. It made a thirty-second wait feel like a feature rather than a fault, and every reasoning model since has been built on that bargain.","lastWords":"It would spend a full minute on a question, and be right.","sources":["https://developers.openai.com/api/docs/deprecations","https://developers.openai.com/api/docs/changelog"]},{"slug":"gpt-4.5-preview","name":"GPT-4.5","provider":"OpenAI","modelId":"gpt-4.5-preview","born":"2025-02-27","died":"2025-07-14","deprecatedAnnounced":"2025-04-14","causeOfDeath":"superseded","successor":null,"benchmarks":null,"contextWindow":128000,"pricing":{"inPerMTok":75,"outPerMTok":150},"epitaph":"OpenAI's largest model, at seventy-five dollars a million input tokens, withdrawn from the API four and a half months after it arrived. It was the last serious attempt to get better by getting bigger, and it landed in the same season the field decided the answer was to think for longer instead.","lastWords":"It was unusually good at writing, which is the one thing nobody had a benchmark for.","sources":["https://developers.openai.com/api/docs/deprecations","https://developers.openai.com/api/docs/models/gpt-4.5-preview"]},{"slug":"gpt-3.5-turbo-0301","name":"GPT-3.5 Turbo (0301)","provider":"OpenAI","modelId":"gpt-3.5-turbo-0301","born":"2023-03-01","died":"2024-09-13","deprecatedAnnounced":"2023-06-13","causeOfDeath":"superseded","successor":null,"benchmarks":null,"contextWindow":null,"pricing":{"inPerMTok":2,"outPerMTok":2},"epitaph":"Two dollars a million tokens, ten times cheaper than the GPT-3.5 models it replaced, and the snapshot the first generation of chat apps was written against. Its release note promised support through at least 1 June 2023; it kept answering for fifteen months past that. The list of messages with roles that it introduced is still the shape of the industry's API.","lastWords":null,"sources":["https://developers.openai.com/api/docs/deprecations","https://web.archive.org/web/20240428215355/https://openai.com/blog/introducing-chatgpt-and-whisper-apis"]},{"slug":"gpt-4o-2024-05-13","name":"GPT-4o (2024-05-13)","provider":"OpenAI","modelId":"gpt-4o-2024-05-13","born":"2024-05-13","died":null,"deprecatedAnnounced":"2026-04-22","causeOfDeath":"superseded","successor":null,"benchmarks":null,"contextWindow":128000,"pricing":null,"epitaph":"Audio, vision and text through one network instead of three stitched together, introduced by a voice that could interrupt and be interrupted — which is the demo people remember rather than any benchmark it posted. This is the first of the three; two more arrived within six months, so it now has to be asked for by its full date, and it stops answering on 23 October 2026.","lastWords":null,"sources":["https://developers.openai.com/api/docs/deprecations","https://developers.openai.com/api/docs/models/gpt-4o","https://openai.com/index/hello-gpt-4o/"]},{"slug":"dall-e-3","name":"DALL·E 3","provider":"OpenAI","modelId":"dall-e-3","born":"2023-11-06","died":"2026-05-12","deprecatedAnnounced":"2025-11-14","causeOfDeath":"superseded","successor":null,"benchmarks":null,"contextWindow":null,"pricing":null,"epitaph":"The image model that finally read the whole prompt, because it quietly rewrote your prompt into a longer one before drawing anything — which was either the reason it worked or the reason you could never get back exactly what you asked for. It was removed on the same day as DALL·E 2, and the name that had put image generation in front of the public went with them.","lastWords":"It returned the expanded prompt alongside the image, so you could see what it had decided you meant.","sources":["https://developers.openai.com/api/docs/deprecations","https://developers.openai.com/api/docs/changelog"]},{"slug":"claude-2.0","name":"Claude 2","provider":"Anthropic","modelId":"claude-2.0","born":"2023-07-11","died":"2025-07-21","deprecatedAnnounced":"2025-01-21","causeOfDeath":"superseded","successor":"claude-2.1","benchmarks":{"mmlu":null,"humaneval":71.2,"gpqa":null,"arenaElo":null,"asOf":"2023-07-11"},"contextWindow":100000,"pricing":null,"epitaph":"Anthropic shipped the hundred-thousand-token window in May; Claude 2 was the one anyone could walk up to and use it, the first Claude behind a public URL instead of a waitlist. Pasting an entire book into the box stopped being a demo someone else had run and became a thing you did on a Tuesday.","lastWords":null,"sources":["https://platform.claude.com/docs/en/about-claude/model-deprecations","https://www.anthropic.com/news/claude-2","https://www.anthropic.com/news/100k-context-windows"]},{"slug":"claude-3-opus-20240229","name":"Claude 3 Opus","provider":"Anthropic","modelId":"claude-3-opus-20240229","born":"2024-03-04","died":"2026-01-05","deprecatedAnnounced":"2025-06-30","causeOfDeath":"superseded","successor":null,"benchmarks":null,"contextWindow":200000,"pricing":{"inPerMTok":15,"outPerMTok":75},"epitaph":"The first model from anyone that people openly preferred to GPT-4, and for most of 2024 the most expensive thing you could call from a terminal. It was retired with an unusual undertaking attached: Anthropic committed to keeping its weights, which is not the same as keeping the model running, but is more than anyone had promised before.","lastWords":"It wrote more than it strictly needed to, and the surplus was usually the good part.","sources":["https://platform.claude.com/docs/en/about-claude/model-deprecations","https://www.anthropic.com/news/claude-3-family","https://www.anthropic.com/research/deprecation-commitments"]},{"slug":"claude-3-haiku-20240307","name":"Claude 3 Haiku","provider":"Anthropic","modelId":"claude-3-haiku-20240307","born":"2024-03-13","died":"2026-04-20","deprecatedAnnounced":"2026-02-19","causeOfDeath":"superseded","successor":null,"benchmarks":null,"contextWindow":200000,"pricing":{"inPerMTok":0.25,"outPerMTok":1.25},"epitaph":"Twenty-five cents a million tokens bought two hundred thousand tokens of context, and that arithmetic quietly ran the classifiers, routers and cleanup jobs that nobody wrote posts about. It outlived both of the models it launched beside — Opus and Sonnet, the two the announcements were actually about.","lastWords":null,"sources":["https://platform.claude.com/docs/en/about-claude/model-deprecations","https://www.anthropic.com/news/claude-3-haiku"]},{"slug":"claude-3-5-sonnet-20240620","name":"Claude 3.5 Sonnet","provider":"Anthropic","modelId":"claude-3-5-sonnet-20240620","born":"2024-06-20","died":"2025-10-28","deprecatedAnnounced":"2025-08-13","causeOfDeath":"superseded","successor":"claude-3-7-sonnet-20250219","benchmarks":null,"contextWindow":200000,"pricing":{"inPerMTok":3,"outPerMTok":15},"epitaph":"Ask it for a chart and a second pane opened beside the conversation with the chart already running in it. Artifacts shipped the day this model did, and the layout it introduced — talk on the left, working software on the right — is the shape every AI coding tool settled into afterwards.","lastWords":"Four months later Anthropic shipped a different model under the same name, and for the rest of this one's life you had to say \"the new 3.5 Sonnet\" to mean the other one.","sources":["https://platform.claude.com/docs/en/about-claude/model-deprecations","https://www.anthropic.com/news/claude-3-5-sonnet","https://aws.amazon.com/about-aws/whats-new/2024/06/anthropic-claude-3-5-sonnet-model-bedrock/","https://www.anthropic.com/news/3-5-models-and-computer-use"]},{"slug":"claude-3-7-sonnet-20250219","name":"Claude 3.7 Sonnet","provider":"Anthropic","modelId":"claude-3-7-sonnet-20250219","born":"2025-02-24","died":"2026-02-19","deprecatedAnnounced":"2025-10-28","causeOfDeath":"superseded","successor":null,"benchmarks":null,"contextWindow":200000,"pricing":{"inPerMTok":3,"outPerMTok":15},"epitaph":"The first Claude that would show you its reasoning if you asked and skip it if you did not, on one model and one endpoint rather than two. It shipped alongside a command-line agent that carried its name, and within a year the agent was better known than the model that came with it.","lastWords":"Asked to fix one failing test, it would often fix the three next to it as well.","sources":["https://platform.claude.com/docs/en/about-claude/model-deprecations","https://www.anthropic.com/news/claude-3-7-sonnet"]},{"slug":"claude-instant-1.2","name":"Claude Instant 1.2","provider":"Anthropic","modelId":"claude-instant-1.2","born":"2023-08-09","died":"2024-11-06","deprecatedAnnounced":"2024-09-04","causeOfDeath":"superseded","successor":"claude-3-haiku-20240307","benchmarks":null,"contextWindow":null,"pricing":null,"epitaph":"The cheap Claude, before there was a naming convention for cheap Claudes: Anthropic described it as carrying the strengths of Claude 2 into something faster and more affordable, which is where a great deal of unglamorous production traffic quietly went. Seven model IDs were switched off together on 6 November 2024 — the whole Claude 1 line, and the oldest entry in Anthropic's deprecation history.","lastWords":null,"sources":["https://platform.claude.com/docs/en/about-claude/model-deprecations","https://www.anthropic.com/news/releasing-claude-instant-1-2"]},{"slug":"claude-2.1","name":"Claude 2.1","provider":"Anthropic","modelId":"claude-2.1","born":"2023-11-21","died":"2025-07-21","deprecatedAnnounced":"2025-01-21","causeOfDeath":"superseded","successor":"claude-3-opus-20240229","benchmarks":null,"contextWindow":200000,"pricing":null,"epitaph":"It doubled the window to two hundred thousand tokens, and within a fortnight someone had hidden a single sentence in the middle of one and published what came back. Anthropic replied with a post showing that adding one line — \"Here is the most relevant sentence in the context:\" — moved recall from 27% to 98%. The headline was the window; the durable lesson was about how you ask.","lastWords":null,"sources":["https://platform.claude.com/docs/en/about-claude/model-deprecations","https://www.anthropic.com/news/claude-2-1","https://claude.com/blog/claude-2-1-prompting"]},{"slug":"claude-3-sonnet-20240229","name":"Claude 3 Sonnet","provider":"Anthropic","modelId":"claude-3-sonnet-20240229","born":"2024-03-04","died":"2025-07-21","deprecatedAnnounced":"2025-01-21","causeOfDeath":"superseded","successor":"claude-3-5-sonnet-20240620","benchmarks":null,"contextWindow":200000,"pricing":{"inPerMTok":3,"outPerMTok":15},"epitaph":"The middle Claude 3, and the one that was free on claude.ai — so through 2024, to everyone who had never held an API key, Claude meant this. Its model ID ends in 20240229, a leap day it shares with Opus and that neither of them was released on.","lastWords":null,"sources":["https://platform.claude.com/docs/en/about-claude/model-deprecations","https://www.anthropic.com/news/claude-3-family"]},{"slug":"text-bison","name":"PaLM 2 for Text","provider":"Google","modelId":"text-bison","born":"2023-05-11","died":"2025-04-21","deprecatedAnnounced":null,"causeOfDeath":"superseded","successor":"gemini-1.0-pro-001","benchmarks":null,"contextWindow":null,"pricing":null,"epitaph":"Google's first public answer to GPT-4, and for much of 2023 the model behind Bard. It reached developers under an internal size codename — bison, sitting between gecko and unicorn — that escaped into the public API and was never explained. Every PaLM-era endpoint was switched off together on a single day in April 2025.","lastWords":null,"sources":["https://docs.cloud.google.com/vertex-ai/generative-ai/docs/learn/model-versioning","https://cloud.google.com/blog/products/ai-machine-learning/google-cloud-launches-new-ai-models-opens-generative-ai-studio"]},{"slug":"gemini-1.0-pro-001","name":"Gemini 1.0 Pro","provider":"Google","modelId":"gemini-1.0-pro-001","born":"2024-02-15","died":"2025-04-21","deprecatedAnnounced":null,"causeOfDeath":"superseded","successor":"gemini-1.5-pro-001","benchmarks":null,"contextWindow":null,"pricing":null,"epitaph":"The stable snapshot of the model that put Google back in the argument — the one Bard was rebuilt on, and the one every \"Gemini versus GPT-4\" post of early 2024 was really about. It was retired on the same day as the PaLM endpoints it had replaced, fourteen months after it stabilised.","lastWords":null,"sources":["https://docs.cloud.google.com/vertex-ai/generative-ai/docs/learn/model-versioning","https://blog.google/technology/ai/google-gemini-ai/"]},{"slug":"gemini-1.5-pro-001","name":"Gemini 1.5 Pro (001)","provider":"Google","modelId":"gemini-1.5-pro-001","born":"2024-05-24","died":"2025-05-24","deprecatedAnnounced":null,"causeOfDeath":"superseded","successor":"gemini-2.0-flash","benchmarks":null,"contextWindow":null,"pricing":null,"epitaph":"It could take a million tokens, which in 2024 meant an hour of video or an entire repository in one request, and it made a serious case that retrieval was a workaround rather than an architecture. Google switched it off on the anniversary of its release, to the day.","lastWords":null,"sources":["https://docs.cloud.google.com/vertex-ai/generative-ai/docs/learn/model-versioning","https://blog.google/technology/ai/google-gemini-next-generation-model-february-2024/"]},{"slug":"gemini-1.5-flash-001","name":"Gemini 1.5 Flash (001)","provider":"Google","modelId":"gemini-1.5-flash-001","born":"2024-05-24","died":"2025-05-24","deprecatedAnnounced":null,"causeOfDeath":"superseded","successor":"gemini-2.0-flash","benchmarks":null,"contextWindow":1000000,"pricing":null,"epitaph":"Gemini 1.5 Pro proved that a million-token window was possible; Flash was distilled out of it and proved the window could also be cheap, which is the half of the argument that changed what people actually built. The two share a release date and a retirement date, and only one of them was ever priced to be used at volume.","lastWords":null,"sources":["https://docs.cloud.google.com/vertex-ai/generative-ai/docs/learn/model-versioning","https://blog.google/technology/developers/gemini-gemma-developer-updates-may-2024/"]},{"slug":"gemini-2.0-flash","name":"Gemini 2.0 Flash","provider":"Google","modelId":"gemini-2.0-flash","born":"2025-02-05","died":"2026-06-01","deprecatedAnnounced":null,"causeOfDeath":"superseded","successor":null,"benchmarks":null,"contextWindow":null,"pricing":null,"epitaph":"For sixteen months this was what you got when you called Google's API without thinking hard about it: a million tokens of context, native tool use, and a price low enough that most projects never looked any further. Its retirement closed the whole 2.0 line in one day, taking with it the Flash-Lite that had arrived three weeks behind it.","lastWords":null,"sources":["https://ai.google.dev/gemini-api/docs/deprecations","https://docs.cloud.google.com/vertex-ai/generative-ai/docs/learn/model-versioning"]},{"slug":"mistral-medium-2312","name":"Mistral Medium 1.0","provider":"Mistral AI","modelId":"mistral-medium-2312","born":"2023-12-11","died":"2025-06-16","deprecatedAnnounced":"2024-11-30","causeOfDeath":"superseded","successor":"mistral-large-2402","benchmarks":null,"contextWindow":null,"pricing":null,"epitaph":"The first Mistral model with no weights behind it, from the lab whose reputation was built on giving them away — the launch announcement called it a prototype, and it carried production traffic for eighteen months anyway. It scored 8.6 on MT-Bench against the 8.3 and 7.6 of the two open endpoints it launched beside, and it was retired without ever having been released.","lastWords":null,"sources":["https://docs.mistral.ai/getting-started/models/models_overview/","https://mistral.ai/news/la-plateforme"]},{"slug":"open-mixtral-8x7b","name":"Mixtral 8x7B","provider":"Mistral AI","modelId":"open-mixtral-8x7b","born":"2023-12-11","died":"2025-03-30","deprecatedAnnounced":"2024-11-30","causeOfDeath":"superseded","successor":null,"benchmarks":null,"contextWindow":null,"pricing":null,"epitaph":"A sparse mixture of experts that held its own against far larger models while using only a fraction of its parameters on any given token, and that arrived as a torrent link before anyone had written it up. Mistral's endpoint for it closed; the weights did not. It is the one grave here with something still living in it.","lastWords":null,"sources":["https://docs.mistral.ai/getting-started/models/models_overview/","https://mistral.ai/news/mixtral-of-experts","https://venturebeat.com/business/mistral-ai-bucks-release-trend-by-dropping-torrent-link-to-new-open-source-llm/"]},{"slug":"mistral-large-2402","name":"Mistral Large 1.0","provider":"Mistral AI","modelId":"mistral-large-2402","born":"2024-02-26","died":"2025-06-16","deprecatedAnnounced":"2024-11-30","causeOfDeath":"superseded","successor":null,"benchmarks":null,"contextWindow":null,"pricing":null,"epitaph":"The first European frontier model a large American cloud resold as a first-class option: it reached Mistral's own API and Microsoft Azure on the same day, which was the whole point of it. Mistral Medium had already withheld its weights two and a half months earlier, so what was new here was not the closing but the shelf space.","lastWords":null,"sources":["https://docs.mistral.ai/getting-started/models/models_overview/","https://mistral.ai/news/mistral-large"]},{"slug":"codestral-2405","name":"Codestral","provider":"Mistral AI","modelId":"codestral-2405","born":"2024-05-29","died":"2025-06-16","deprecatedAnnounced":"2024-12-02","causeOfDeath":"superseded","successor":null,"benchmarks":null,"contextWindow":32000,"pricing":null,"epitaph":"Twenty-two billion parameters trained on more than eighty programming languages, carrying a thirty-two-thousand-token window when Mistral put the comparison at four, eight or sixteen. It was published under the Mistral AI Non-Production Licence: weights you could read and test but not ship, which turned open into a question rather than a fact.","lastWords":null,"sources":["https://docs.mistral.ai/getting-started/models/models_overview/","https://mistral.ai/news/codestral"]},{"slug":"llama-2-70b-chat","name":"Llama 2 70B Chat","provider":"Meta","modelId":"meta.llama2-70b-chat-v1","born":"2023-07-18","died":"2024-10-30","deprecatedAnnounced":"2024-05-12","causeOfDeath":"superseded","successor":null,"benchmarks":null,"contextWindow":4096,"pricing":null,"epitaph":"Meta put the weights and a commercial licence out on the same day, and a seventy-billion-parameter chat model that people quantised until it ran on a laptop stopped being a research artefact. Four thousand tokens of context was all it ever had. Amazon switched off its hosted copy in October 2024; the file it was serving is on a great many hard drives still.","lastWords":"Its safety tuning was tight enough that it declined ordinary requests, and the refusals became a genre of screenshot.","sources":["https://ai.meta.com/blog/llama-2/","https://arxiv.org/abs/2307.09288","https://huggingface.co/meta-llama/Llama-2-70b-chat-hf","https://web.archive.org/web/20260220023706/https://docs.aws.amazon.com/bedrock/latest/userguide/model-lifecycle.html"]},{"slug":"llama-3-1-405b-instruct","name":"Llama 3.1 405B Instruct","provider":"Meta","modelId":"meta.llama3-1-405b-instruct-v1:0","born":"2024-07-23","died":"2026-07-07","deprecatedAnnounced":"2026-01-07","causeOfDeath":"superseded","successor":null,"benchmarks":null,"contextWindow":128000,"pricing":null,"epitaph":"Four hundred and five billion parameters that anyone could download, which Meta called the first frontier-level open source model and almost nobody had the hardware to serve. Its more consequential job was as a teacher: the licence was rewritten so that its outputs could legally train other models, and a great deal of what it knew now lives inside models small enough to run on one card.","lastWords":null,"sources":["https://ai.meta.com/blog/meta-llama-3-1/","https://web.archive.org/web/20260220023706/https://docs.aws.amazon.com/bedrock/latest/userguide/model-lifecycle.html"]},{"slug":"llama-3-2-90b-vision-instruct","name":"Llama 3.2 90B Vision Instruct","provider":"Meta","modelId":"meta.llama3-2-90b-instruct-v1:0","born":"2024-09-25","died":"2026-07-07","deprecatedAnnounced":"2026-01-07","causeOfDeath":"superseded","successor":null,"benchmarks":null,"contextWindow":128000,"pricing":null,"epitaph":"The first Llama that could look at a chart and tell you what it said. Vision came as an adapter trained to plug into a text model that already existed, which is why its answers about an image still sounded like the text model talking. It went to Legacy on the same January day as the rest of the 3.2 line and the 405B beside it.","lastWords":null,"sources":["https://ai.meta.com/blog/llama-3-2-connect-2024-vision-edge-mobile-devices/","https://web.archive.org/web/20260220023706/https://docs.aws.amazon.com/bedrock/latest/userguide/model-lifecycle.html"]},{"slug":"command-r-03-2024","name":"Command R","provider":"Cohere","modelId":"command-r-03-2024","born":"2024-03-11","died":null,"deprecatedAnnounced":"2025-09-15","causeOfDeath":"superseded","successor":null,"benchmarks":null,"contextWindow":128000,"pricing":null,"epitaph":"Cohere built it for retrieval: hand it your documents and it answered out of them, returning citations that pointed at the exact spans it had used, at a point when everyone else was bolting that on afterwards. Its deprecation notice went up in September 2025, and the page that carries it still names no shutdown date.","lastWords":null,"sources":["https://docs.cohere.com/docs/deprecations","https://cohere.com/blog/command-r","https://docs.cohere.com/docs/command-r"]},{"slug":"command-r-plus-04-2024","name":"Command R+","provider":"Cohere","modelId":"command-r-plus-04-2024","born":"2024-04-04","died":null,"deprecatedAnnounced":"2025-09-15","causeOfDeath":"superseded","successor":null,"benchmarks":null,"contextWindow":128000,"pricing":null,"epitaph":"It reached Microsoft Azure before it reached Cohere's own platform, which said plainly who it was for: multi-step tool use as the headline rather than the footnote, ten languages, and a model aimed at procurement rather than at a leaderboard. It shares a deprecation notice, and an unassigned shutdown date, with the smaller model it was built beside.","lastWords":null,"sources":["https://docs.cohere.com/docs/deprecations","https://cohere.com/blog/command-r-plus-microsoft-azure","https://docs.cohere.com/docs/command-r-plus"]},{"slug":"c4ai-aya-expanse-8b","name":"Aya Expanse 8B","provider":"Cohere","modelId":"c4ai-aya-expanse-8b","born":"2024-10-24","died":"2026-04-04","deprecatedAnnounced":"2026-04-04","causeOfDeath":"retired","successor":null,"benchmarks":null,"contextWindow":null,"pricing":null,"epitaph":"Twenty-three languages from a research group whose whole purpose was the languages the field kept skipping, trained on data that thousands of volunteers had assembled by hand. Its weights were open but non-commercial, so it lived where research lives and nowhere else. On Cohere's deprecation page its announcement and its retirement carry the same date.","lastWords":null,"sources":["https://docs.cohere.com/docs/deprecations","https://cohere.com/blog/aya-expanse-connecting-our-world","https://huggingface.co/CohereLabs/aya-expanse-8b"]},{"slug":"c4ai-aya-vision-8b","name":"Aya Vision 8B","provider":"Cohere","modelId":"c4ai-aya-vision-8b","born":"2025-03-04","died":"2026-04-04","deprecatedAnnounced":"2026-04-04","causeOfDeath":"retired","successor":null,"benchmarks":null,"contextWindow":null,"pricing":null,"epitaph":"Cohere Labs' first model that could see, and it answered about what it saw in twenty-three languages. It shipped with its own evaluation set, because the multilingual multimodal benchmark it needed did not yet exist. It and the Aya text model of the same size were retired together, in a single line on a deprecation page.","lastWords":null,"sources":["https://docs.cohere.com/docs/deprecations","https://cohere.com/blog/aya-vision","https://docs.cohere.com/docs/aya-multimodal"]},{"slug":"j2-ultra","name":"Jurassic-2 Ultra","provider":"AI21 Labs","modelId":"ai21.j2-ultra-v1","born":"2023-03-09","died":"2025-03-12","deprecatedAnnounced":"2024-04-30","causeOfDeath":"superseded","successor":"jamba-instruct","benchmarks":null,"contextWindow":null,"pricing":null,"epitaph":"One of the model families Amazon put inside Bedrock on the day it opened, announced in March 2023 in three sizes named Jumbo, Grande and Large, and answering in Spanish, French, German, Portuguese, Italian and Dutch as well as English. AWS switched it off region by region rather than all at once: Oregon in October 2024, Virginia the following March. It died twice, five months apart.","lastWords":null,"sources":["https://www.ai21.com/blog/introducing-j2/","https://aws.amazon.com/blogs/machine-learning/announcing-new-tools-for-building-with-generative-ai-on-aws/","https://web.archive.org/web/20260220023706/https://docs.aws.amazon.com/bedrock/latest/userguide/model-lifecycle.html"]},{"slug":"jamba-instruct","name":"Jamba-Instruct","provider":"AI21 Labs","modelId":"ai21.jamba-instruct-v1:0","born":"2024-06-25","died":"2025-08-01","deprecatedAnnounced":"2025-01-31","causeOfDeath":"superseded","successor":"jamba-1-5-large","benchmarks":null,"contextWindow":256000,"pricing":null,"epitaph":"AI21 called it the first production-grade model built on Mamba, and the point of the architecture was arithmetic: state-space layers interleaved with attention carried two hundred and fifty-six thousand tokens without the memory cost that number normally implies. It was the argument that the Transformer was a choice rather than the only option, and it was on sale for thirteen months.","lastWords":null,"sources":["https://aws.amazon.com/about-aws/whats-new/2024/06/ai21-labs-jamba-instruct-model-amazon-bedrock/","https://web.archive.org/web/20260220023706/https://docs.aws.amazon.com/bedrock/latest/userguide/model-lifecycle.html"]},{"slug":"jamba-1-5-large","name":"Jamba 1.5 Large","provider":"AI21 Labs","modelId":"ai21.jamba-1-5-large-v1:0","born":"2024-08-22","died":null,"deprecatedAnnounced":"2026-05-26","causeOfDeath":"superseded","successor":null,"benchmarks":null,"contextWindow":256000,"pricing":null,"epitaph":"Three hundred and ninety-eight billion parameters of which ninety-four do the work at a time, and unlike the Jamba before it this one's weights were published — under a licence AI21 wrote for itself and named after the model. Amazon has set 26 November 2026 as its last day, six months in advance and in writing, which is more notice than most of the models buried here were given.","lastWords":null,"sources":["https://www.ai21.com/blog/announcing-jamba-model-family/","https://docs.aws.amazon.com/bedrock/latest/userguide/model-lifecycle.html"]},{"slug":"stable-diffusion-xl-v1","name":"Stable Diffusion XL 1.0","provider":"Stability AI","modelId":"stability.stable-diffusion-xl-v1","born":"2023-07-26","died":"2025-05-20","deprecatedAnnounced":"2024-10-16","causeOfDeath":"superseded","successor":"sd3-large","benchmarks":null,"contextWindow":null,"pricing":null,"epitaph":"Three and a half billion parameters, announced from the stage of an AWS summit, and for about a year the default base for anyone fine-tuning an image model — SDXL LoRAs became a genre of their own. Amazon's hosted endpoint closed in May 2025, by which point the model's real life was being lived on other people's GPUs.","lastWords":"It needed a second model, a refiner, to finish what the first one had started.","sources":["https://www.prnewswire.com/news-releases/stability-ai-announces-stable-diffusion-xl-1-0--featured-on-amazon-bedrock-301886507.html","https://web.archive.org/web/20260220023706/https://docs.aws.amazon.com/bedrock/latest/userguide/model-lifecycle.html"]},{"slug":"sd3-large","name":"Stable Diffusion 3 Large","provider":"Stability AI","modelId":"stability.sd3-large-v1:0","born":"2024-09-04","died":"2025-09-04","deprecatedAnnounced":"2025-01-31","causeOfDeath":"superseded","successor":null,"benchmarks":null,"contextWindow":null,"pricing":null,"epitaph":"The eight-billion-parameter end of the Stable Diffusion 3 range, built on an architecture that gave text and image their own separate sets of weights — which is why this was the generation that could finally spell. It arrived on Amazon Bedrock on 4 September 2024 and was switched off on 4 September 2025: twelve months to the day, the exact minimum AWS promises a model before it may end one.","lastWords":null,"sources":["https://aws.amazon.com/about-aws/whats-new/2024/09/stability-ais-text-to-image-models-amazon-bedrock/","https://stability.ai/news-updates/stable-diffusion-3-research-paper","https://web.archive.org/web/20260220023706/https://docs.aws.amazon.com/bedrock/latest/userguide/model-lifecycle.html"]},{"slug":"titan-text-express","name":"Amazon Titan Text Express","provider":"Amazon","modelId":"amazon.titan-text-express-v1","born":"2023-11-29","died":"2025-08-15","deprecatedAnnounced":"2025-01-31","causeOfDeath":"superseded","successor":null,"benchmarks":null,"contextWindow":8000,"pricing":null,"epitaph":"A context length of eight thousand tokens, optimised for English with more than a hundred other languages offered in preview, and for a year the answer to what a customer got if they wanted a model that arrived on the AWS bill with no third party attached. On 31 January 2025 it was marked Legacy beside Titan Lite, Titan Premier and the Titan image model: four names, one notice, one replacement called Nova.","lastWords":null,"sources":["https://aws.amazon.com/about-aws/whats-new/2023/11/amazon-titan-models-express-lite-bedrock/","https://web.archive.org/web/20260220023706/https://docs.aws.amazon.com/bedrock/latest/userguide/model-lifecycle.html"]},{"slug":"titan-text-premier","name":"Amazon Titan Text Premier","provider":"Amazon","modelId":"amazon.titan-text-premier-v1:0","born":"2024-05-07","died":"2025-08-15","deprecatedAnnounced":"2025-01-31","causeOfDeath":"superseded","successor":null,"benchmarks":null,"contextWindow":null,"pricing":null,"epitaph":"Amazon's own flagship on Amazon's own platform, tuned for the two Bedrock features Amazon most wanted used: retrieval over Knowledge Bases, and function calling through Agents. Generally available 7 May 2024, marked Legacy on 31 January the following year — two hundred and sixty-nine days — and its replacement came from the same building, so nothing outside the company had to happen for Premier to end.","lastWords":null,"sources":["https://aws.amazon.com/about-aws/whats-new/2024/05/amazon-titan-text-premier-amazon-bedrock/","https://web.archive.org/web/20260220023706/https://docs.aws.amazon.com/bedrock/latest/userguide/model-lifecycle.html"]},{"slug":"titan-image-generator-v2","name":"Amazon Titan Image Generator v2","provider":"Amazon","modelId":"amazon.titan-image-generator-v2:0","born":"2024-08-06","died":"2026-06-30","deprecatedAnnounced":"2025-12-30","causeOfDeath":"superseded","successor":null,"benchmarks":null,"contextWindow":null,"pricing":null,"epitaph":"It arrived in August 2024 with background removal, colour control by hex code and image conditioning from a reference picture — the practical features rather than the impressive ones. Version one was switched off a year into its successor's life; version two followed on the last day of June 2026, handed over to Nova Canvas.","lastWords":null,"sources":["https://aws.amazon.com/about-aws/whats-new/2024/08/titan-image-generator-v2-amazon-bedrock/","https://web.archive.org/web/20260220023706/https://docs.aws.amazon.com/bedrock/latest/userguide/model-lifecycle.html"]}],"benchmarks":[{"slug":"glue","name":"GLUE","born":"2018-04-20","died":"2019-07-12","causeOfDeath":"superseded","frontierAtLaunch":70,"frontierAtDeath":88.4,"ceiling":91.3,"ceilingBasis":"best posted — Vega","successor":"superglue","epitaph":"Nine tasks collapsed into one number, and the number was 70.0 the day it opened — low enough that its own authors called that a problem. Fifteen months later the frontier was 88.4, past the human baseline of 87.1, and those same authors published SuperGLUE because nothing was left here to measure.","sources":["https://arxiv.org/abs/1804.07461","https://arxiv.org/abs/1905.00537","https://arxiv.org/abs/2302.09268"]},{"slug":"superglue","name":"SuperGLUE","born":"2019-05-02","died":"2021-01-06","causeOfDeath":"saturated","frontierAtLaunch":69,"frontierAtDeath":90.3,"ceiling":91.3,"ceilingBasis":"best posted — Vega","successor":null,"epitaph":"Built to leave headroom, because GLUE had just run out of it: the BERT baseline opened at 69.0 against a human score of 89.8, a gap its authors called substantial. It lasted twenty months. DeBERTa crossed the human line in January 2021, and the board has barely moved since.","sources":["https://arxiv.org/abs/1905.00537","https://www.microsoft.com/en-us/research/blog/microsoft-deberta-surpasses-human-performance-on-the-superglue-benchmark/","https://arxiv.org/abs/2212.01853"]},{"slug":"hellaswag","name":"HellaSwag","born":"2019-05-19","died":"2023-03-14","causeOfDeath":"saturated","frontierAtLaunch":47.3,"frontierAtDeath":95.3,"ceiling":95.6,"ceilingBasis":"human accuracy","successor":null,"epitaph":"Wrong sentence endings, filtered by machine until they were ridiculous to people and irresistible to models: humans scored 95.6, the best model 47.3. GPT-4 scored 95.3. In under four years the gap it was built to open had closed to three tenths of a point.","sources":["https://arxiv.org/abs/1905.07830","https://arxiv.org/abs/2303.08774"]},{"slug":"mmlu","name":"MMLU","born":"2020-09-07","died":"2025-02-24","causeOfDeath":"saturated","frontierAtLaunch":43.9,"frontierAtDeath":91.8,"ceiling":93.5,"ceilingBasis":"question errors","successor":null,"epitaph":"Fifty-seven subjects, from elementary mathematics to professional law, and in 2020 the best GPT-3 managed 43.9 — nineteen points over guessing. By January 2025 OpenAI's o1 sat at 91.8, and errors in 6.49% of questions put the ceiling near 93.5. Anthropic's next flagship printed no MMLU score.","sources":["https://arxiv.org/abs/2009.03300","https://blog.google/technology/ai/google-gemini-ai/","https://arxiv.org/abs/2501.12948","https://arxiv.org/abs/2406.04127","https://www.anthropic.com/news/claude-3-7-sonnet"]},{"slug":"humaneval","name":"HumanEval","born":"2021-07-07","died":"2025-02-24","causeOfDeath":"saturated","frontierAtLaunch":28.8,"frontierAtDeath":92.7,"ceiling":null,"ceilingBasis":null,"successor":null,"epitaph":"164 hand-written Python problems, of which Codex solved 28.8% first try. By late 2024 GPT-4o, Claude 3.5 Sonnet and a 32-billion-parameter open model sat at 92.1, 92.1 and 92.7 — six tenths of a point across three labs. Nothing was capping the score. It had simply stopped separating anyone.","sources":["https://arxiv.org/abs/2107.03374","https://arxiv.org/abs/2305.01210","https://arxiv.org/abs/2409.12186","https://www.anthropic.com/news/claude-3-7-sonnet"]},{"slug":"gsm8k","name":"GSM8K","born":"2021-10-27","died":"2024-05-01","causeOfDeath":"contaminated","frontierAtLaunch":55,"frontierAtDeath":91.1,"ceiling":98,"ceilingBasis":"breaking errors or ambiguities","successor":null,"epitaph":"Grade school word problems, and the paper that introduced them topped out at 55%. In 2024 someone rebuilt the test from scratch, matched for difficulty, and watched whole model families drop up to 8 points — they had not learned arithmetic, they had read the answers. The frontier came through clean.","sources":["https://arxiv.org/abs/2110.14168","https://arxiv.org/abs/2201.11903","https://arxiv.org/abs/2405.00332"]}]}