Lifespans
Half the models the museum has watched were retired within 439 days of their release, about 14 months.
95% interval, 399 to 511 days. 267 models, 70 retirements, watched from 20 Feb 2026 to 23 Oct 2026.
- Gone within a year
- 38%29% to 48%
- Gone within two years
- 83%71% to 92%
Next to go
Retirements the labs have announced for models OpenRouter still serves. The dates are for each lab’s own API.
- Claude Sonnet 4.5Anthropic, announced 30 Sep 2026
- GPT-5, GPT-5 Mini, GPT-5 Nano, GPT-5 Pro, OpenAI o3OpenAI, announced 11 Jun 2026
- GPT-4o (Omni), GPT-4o mini, GPT-4.1, GPT-4.1 MiniOpenAI, announced 14 Oct 2026
- GPT-5.1, GPT-5.3-Codex, GPT-5.4 NanoOpenAI, announced 1 Oct 2026
- Mistral Large 3 2512Mistral, announced 31 Oct 2026
- Gemini 3.1 Flash LiteGoogle
The record
Every retirement the museum has watched, newest first: the last day the model was served, and how long it lived.
- Solar Pro 3275 days
- GPT-4 Turbo932 days
- GPT-5.2 Chat322 days
- o1692 days
- o3 Mini682 days
- OpenAI o4 Mini High560 days
- OpenAI o4-mini560 days
- Qwen3 Next 80B A3B Thinking412 days
- NVIDIA Nemotron 3 Super230 days
- GPT-3.5 Turbo1,333 days
- DeepSeek R1625 days
- Claude Sonnet 4511 days
58 earlier retirements
- Qwen Plus 0728396 days
- Qwen3 235B A22B529 days
- Qwen3 30B A3B Thinking 2507406 days
- Qwen3 Coder Plus387 days
- Qwen3 Max399 days
- Qwen3 Max Thinking242 days
- Qwen3.6 Max Preview165 days
- Ling 3.0 Flash Sante34 days
- Ling 3.0 Flash Fin32 days
- Claude 3 Haiku935 days
- Nex-N2.5-Mini17 days
- Nex-N2.5-Pro17 days
- Ling 3.0 Flash VL13 days
- Claude Opus 4487 days
- Gemini 2.5 Pro (I/O Edition)496 days
- Gemini 2.5 Pro Preview 06-05466 days
- Mercury 2.5 Preview8 days
- Qwen3.8 Max32 days
- GPT-5 Codex338 days
- Gemma 3n 4B462 days
- NVIDIA Nemotron Nano 9B V2353 days
- GPT-5.3 Chat160 days
- GPT-5.1 Chat260 days
- Gemini 2.5 Flash Lite Preview 09-2025287 days
- INTELLECT-3208 days
- Llama 3 70B792 days
- MiMo-V2-Flash185 days
- GLM 4 32B320 days
- ERNIE 4.5 300B A47B337 days
- MiMo-V2-Omni74 days
- MiMo-V2-Pro74 days
- Mistral Devstral Medium324 days
- Mistral Devstral Small 1.1324 days
- GPT-4o Audio435 days
- DeepSeek V3.2 Speciale178 days
- Trinity Large Preview479 days
- Grok 3451 days
- Grok 3 Beta401 days
- Grok 3 Mini Beta401 days
- Grok 4310 days
- Grok 4 Fast238 days
- Grok 4.1 Fast175 days
- Grok Code Fast 1262 days
- Ring 2.6 1T7 days
- Claude 3.7 Sonnet439 days
- Claude 3.7 Thinking Sonnet438 days
- Ling 2.6 1T14 days
- Gemma 3n 2B300 days
- Mistral Small Creative134 days
- Ling 2.6 Flash7 days
- QwQ 32B419 days
- Mercury292 days
- Claude 3.5 Sonnet653 days
- Qwen3.6 Plus Preview3 days
- Grok 4.20 Beta19 days
- Grok 4.20 Multi-Agent Beta19 days
- Gemini 3 Pro Preview127 days
- Llama 3.1 405B601 days
How this is counted
A model’s life runs from its release to the last day it was served on OpenRouter. The museum’s watch checks OpenRouter every day and calls a model retired once none of its listings has been served for 36 hours. A model still served counts as alive up to 23 Oct 2026, when these statuses were taken; in survival analysis it is censored there.
Retirements are dated only from 20 Feb 2026, when OpenRouter availability began to be recorded, and a model enters the collection only once it has been asked. So each model counts from the later of those two days, at the age it had then (late entry, or left truncation). The 26 models already gone by then are left out rather than dated to that day. Read without this correction, the same records give a median of 560 days.
Also left out: 19 code names, the previews OpenRouter serves under invented names (counting them gives a median of 435 days), and 70 models OpenRouter never listed, whose availability is not observed. Served means served on OpenRouter, which can differ from a lab’s own API.
The estimate is Kaplan and Meier’s, with Greenwood’s variance and a log-log 95% interval; the median’s interval is where the interval’s bounds cross one half. Release days are those in the open collection. Nothing here ranks models or labs: a long life says nothing about a model’s answers.
Downloads
- The curve, and one for each lab with ten models or more (CSV)
- Every model counted, with its ages (CSV)
- All of it (JSON)
The numbers are recomputed from the collection whenever a status changes. To cite them, cite the collection:
Museum of Models. (2026). Museum of Models collection (Version 1.6.0) [Data set]. https://doi.org/10.5281/zenodo.23828158