Analytics
280 tracked releases · 27 labs · 2022–2026
Releases per year
How many tracked releases landed each year. 2026 is still in progress.
Releases per lab
Tracked releases by organisation, most active first.
Open weights over time
Open weights Proprietary
Median days between releases
Typical gap between consecutive releases from the same lab. Only labs with three or more tracked releases appear, since a median needs at least two gaps.
Largest context window by year
The biggest disclosed context window among that year's tracked releases. Bars are on a log scale — the range spans three orders of magnitude, so a linear axis would render the early years invisible. Read the labels, not the widths.
Read the full study: how the context window grew, step by step →
What the labs charge, and why the figures are not comparable →
Open-weights licences
How the 113 open-weights releases are licensed. Permissive licences (Apache-2.0, MIT) sit alongside bespoke community terms that carry their own restrictions — check the licence before assuming reuse is free.
Capabilities
How often each capability is tagged across all tracked releases.
When models stopped being text-only
Share of each year's releases that accept or produce more than text, counting only releases whose modalities have been researched. The denominator is small in early years, so read the counts rather than the percentages. 163 of 280 records have modalities recorded — see data quality.
Input modalities in use
Across every release with researched modalities. A model counts once per modality it accepts.
When each capability was first evidenced
The earliest tracked release this dataset can SHOW carrying each capability, with the source behind it on the model's page. Not a claim about who did it first — only about what has been evidenced here, from the 163 records whose capabilities have been researched.
- coding PaLM
- multilingual PaLM
- vision GPT-4
- audio Gemini 1.0 Ultra
- video Gemini 1.5 Pro
- structured output Claude 3 Opus
- agentic Command R
- function calling Command R
- tool use Command R
- reasoning o1
- image generation Nova Canvas
- video generation Nova Reel
- speech generation AuK
Capability adoption over time
Releases evidencing each capability, over the releases of that year whose capabilities have been researched. Read the counts: the denominator is small in early years, and a capability absent from a record means it has not been evidenced, never that the model lacks it — see data quality.
| Capability | 2022 | 2023 | 2024 | 2025 | 2026 |
|---|---|---|---|---|---|
| reasoning | 0/4 | 0/13 | 1/25 | 16/36 | 42/85 |
| agentic | 0/4 | 0/13 | 3/25 | 12/36 | 35/85 |
| coding | 1/4 | 6/13 | 7/25 | 12/36 | 16/85 |
| vision | 0/4 | 4/13 | 8/25 | 15/36 | 15/85 |
| function calling | 0/4 | 0/13 | 4/25 | 10/36 | 24/85 |
| multilingual | 3/4 | 1/13 | 4/25 | 8/36 | 10/85 |
| structured output | 0/4 | 0/13 | 3/25 | 4/36 | 19/85 |
| tool use | 0/4 | 0/13 | 4/25 | 4/36 | 2/85 |
| image generation | 0/4 | 0/13 | 1/25 | 0/36 | 4/85 |
| audio | 0/4 | 3/13 | 1/25 | 0/36 | 0/85 |
| video | 0/4 | 0/13 | 3/25 | 0/36 | 1/85 |
Which labs evidence which capabilities
Counted over each lab's researched releases, not all of them, and the denominator is in every cell. A zero means the capability has not been evidenced on that lab's records — never that its models lack it. Labs with fewer than five researched releases are excluded, because at that size a row measures our research rather than the lab.
| Lab | reasoning | agentic | coding | vision | function calling | multilingual | structured output | tool use | image generation | audio | video |
|---|---|---|---|---|---|---|---|---|---|---|---|
| Google DeepMind 30 researched | 18/30 | 10/30 | 5/30 | 7/30 | 16/30 | 2/30 | 12/30 | 0/30 | 4/30 | 3/30 | 2/30 |
| OpenAI 24 researched | 16/24 | 7/24 | 4/24 | 7/24 | 11/24 | 0/24 | 8/24 | 2/24 | 0/24 | 1/24 | 0/24 |
| Alibaba Qwen 21 researched | 7/21 | 5/21 | 5/21 | 0/21 | 0/21 | 2/21 | 0/21 | 0/21 | 0/21 | 0/21 | 0/21 |
| Anthropic 16 researched | 4/16 | 10/16 | 4/16 | 14/16 | 0/16 | 10/16 | 1/16 | 0/16 | 0/16 | 0/16 | 0/16 |
| Zhipu AI 11 researched | 1/11 | 3/11 | 3/11 | 0/11 | 0/11 | 1/11 | 0/11 | 0/11 | 0/11 | 0/11 | 0/11 |
| Meta AI 10 researched | 2/10 | 2/10 | 2/10 | 5/10 | 0/10 | 2/10 | 0/10 | 3/10 | 0/10 | 0/10 | 0/10 |
| Mistral AI 10 researched | 1/10 | 2/10 | 4/10 | 5/10 | 2/10 | 6/10 | 0/10 | 1/10 | 0/10 | 0/10 | 0/10 |
| DeepSeek 8 researched | 3/8 | 1/8 | 3/8 | 0/8 | 2/8 | 0/8 | 1/8 | 1/8 | 0/8 | 0/8 | 0/8 |
| xAI 7 researched | 5/7 | 4/7 | 6/7 | 1/7 | 4/7 | 0/7 | 4/7 | 0/7 | 0/7 | 0/7 | 0/7 |
| Amazon 5 researched | 0/5 | 1/5 | 1/5 | 2/5 | 0/5 | 0/5 | 0/5 | 1/5 | 1/5 | 0/5 | 2/5 |
9 labs with fewer than five researched releases are left out — at one or two records a row says more about how much has been researched than about the lab.
Compare models side by side →
Follow a family through its generations →