Releases and independent scores retrieved from Epoch AI (Data on AI models; Capabilities & benchmarking) on 2026-08-24; record updated 2026-10-19
From ChatGPT to 2026-10-19: 352 releases from twelve tracked labs, one mark each.
Pooled across the twelve labs, the median wait between one release and the next fell from 3.5 days in 2023 to 2 in 2025.
Revisions, updates to a model already on record, are left out of these counts: 14 so far.
Model cards: 29 of 45 checked releases in 2023 (64%), 50 of 93 in 2026 so far (54%). Usage policies: 38 of 45 (84%), then 93 of 93 (100%).
On a second benchmark, SWE-bench Verified (fixing real GitHub issues), Epoch's top score rose from 31 in 2024 to 84 in 2026.
The highest open-weights GPQA Diamond score on record is 91.9 (GLM-5.2, 2026).
In three (Gemini 3.1 Flash-Lite, GPT-5.4 and GPT-5.5) the lab ran a higher reasoning setting than Epoch, so those gaps include a settings difference.
Six pairs from recent releases cannot test whether labs overstate.
For example, median days between releases at first and lately: OpenAI 66 to 22, Anthropic 104 to 21, Google DeepMind 41 to 15, Alibaba (Qwen) 20.5 to 11.5. "At first" is the median of a lab's first third of gaps on record; "lately", of its last third.
ByteDance, Tencent and Baidu are not tracked, so the record cannot support a comparison between countries.
Every view on this page, in words and tables. Every entry, with its sources, is in the explorer.
The first full year after ChatGPT. 11 of the twelve tracked labs released at least one model. Median time between one release and the next: 3.5 days. The busiest week brought 5 releases.
All twelve tracked labs released at least one model. First on record this year: Moonshot AI. Median time between one release and the next: 2 days. The busiest week brought 9 releases.
All twelve tracked labs released at least one model. Median time between one release and the next: 2 days. The busiest week brought 6 releases.
A partial year: 93 releases by 2026-10-19. By the same date in 2025 there were 98. 11 of the twelve tracked labs released at least one model. Median time between one release and the next: 1.5 days. The busiest week brought 7 releases.
| Year | Releases | Revisions |
|---|---|---|
| 2022 | 2 | 0 |
| 2023 | 45 | 1 |
| 2024 | 92 | 3 |
| 2025 | 120 | 7 |
| 2026 | 93 | 3 |
About 66 days between releases at first, about 22 lately.
About 104 days between releases at first, about 21 lately.
About 41 days between releases at first, about 15 lately.
About 20.5 days between releases at first, about 11.5 lately.
| Lab | System card | Model card | Usage policy | Paper |
|---|---|---|---|---|
| OpenAI | 36 / 12 / 0 | 1 / 47 / 0 | 48 / 0 / 0 | 7 / 41 / 0 |
| Anthropic | 18 / 8 / 0 | 5 / 21 / 0 | 26 / 0 / 0 | 0 / 26 / 0 |
| Google DeepMind | 4 / 51 / 0 | 45 / 10 / 0 | 49 / 6 / 0 | 34 / 21 / 0 |
| Meta | 2 / 24 / 0 | 20 / 5 / 1 | 20 / 5 / 1 | 16 / 10 / 0 |
| xAI | 9 / 7 / 2 | 2 / 16 / 0 | 18 / 0 / 0 | 0 / 18 / 0 |
| Mistral | 0 / 41 / 0 | 28 / 13 / 0 | 41 / 0 / 0 | 9 / 32 / 0 |
| DeepSeek | 0 / 25 / 0 | 24 / 1 / 0 | 25 / 0 / 0 | 20 / 5 / 0 |
| Alibaba (Qwen) | 0 / 63 / 0 | 48 / 15 / 0 | 63 / 0 / 0 | 32 / 30 / 1 |
| Moonshot AI | 0 / 10 / 0 | 8 / 2 / 0 | 10 / 0 / 0 | 6 / 4 / 0 |
| Z.ai (Zhipu) | 0 / 16 / 0 | 13 / 3 / 0 | 16 / 0 / 0 | 14 / 2 / 0 |
| Cohere | 2 / 15 / 0 | 13 / 4 / 0 | 17 / 0 / 0 | 6 / 11 / 0 |
| Amazon | 6 / 1 / 0 | 6 / 1 / 0 | 7 / 0 / 0 | 3 / 4 / 0 |
The first 7 independent scores on record; the highest is 36 on GPQA Diamond.
23 scored this year; the highest is 74.
26 scored this year; the highest is 93.
34 scored so far; the highest is 95.
First independent GPQA Diamond score on record: 35.7 (GPT-4, 2023). Latest: 89.9 (GPT-5.6 Sol, 2026). Highest: 90.7 (GPT-5.5). 1 of these 17 scored releases has open weights.
First independent GPQA Diamond score on record: 33.0 (Claude 2.1, 2023). Latest: 92.9 (Claude Opus 5, 2026).
First independent GPQA Diamond score on record: 34.0 (Gemini 1.0 Pro, 2023). Latest: 94.8 (Gemini 3.7 Flash, 2026). 2 of these 12 scored releases have open weights.
First independent GPQA Diamond score on record: 28.8 (Qwen1.5-72B, 2024). Latest: 82.3 (Qwen3.7 Flash, 2026). Highest: 92.7 (Qwen 3.8 Max). 8 of these 14 scored releases have open weights.
On a second benchmark, SWE-bench Verified (fixing real GitHub issues), Epoch's top score rose from 31 in 2024 to 84 in 2026.
On SWE-bench Verified, OpenAI's independent score moves from 48 to 81 across 6 releases on record.
On SWE-bench Verified, Anthropic's independent score moves from 61 to 84 across 8 releases on record.
On SWE-bench Verified, Google DeepMind's independent score moves from 58 to 79 across 3 releases on record.
5 releases were graded twice on the same benchmark: once by the releasing lab, once by Epoch. 2 of the 6 pairs sit within a single point. In three (Gemini 3.1 Flash-Lite, GPT-5.4 and GPT-5.5) the lab ran a higher reasoning setting than Epoch, so those gaps include a settings difference. With so few pairs this is not a test of whether labs overstate. Both numbers are kept; neither is averaged.
| Release | Benchmark | LAB | IND | Lab setting | Epoch setting |
|---|---|---|---|---|---|
| Gemini 3.1 Flash-Lite | GPQA Diamond | 86.9 | 74.2 | high thinking | low effort |
| GPT-5.4 | GPQA Diamond | 92.8 | 84.8 | xhigh effort | low effort |
| GPT-5.5 | GPQA Diamond | 93.6 | 90.7 | xhigh effort | low effort |
| DeepSeek-V4-Pro | GPQA Diamond | 90.1 | 89.6 | max thinking | max effort |
| DeepSeek-V4-Pro | SWE-bench Verified | 80.6 | 77.6 | max thinking | max effort |
| GLM-5.2 | GPQA Diamond | 91.2 | 91.9 | not stated | max effort |
In 2023, 27 of 45 checked releases had a paper on record (60%). In 2026 so far, 22 of 92 (24%). Papers can come out weeks after a release, so recent years may rise.
In 2023, 26 of 45 releases had a parameter count on record (58%). In 2026 so far, 34 of 93 (37%). Counts reported by others are left out. Epoch often adds counts for recent models late, and we have not checked the blanks against the labs' own pages.
From 2023 to 2026, the highest independent GPQA Diamond score rose from 36 to 95. The share of checked releases with a paper fell from 60% to 24%, and the share with a parameter count on record fell from 58% to 37%. The share with a system card rose from 11% to 35% (5 of 45, then 32 of 92).
Papers can come out weeks after a release, so recent years may rise. Parameter counts are Epoch AI's and divide by all releases, leaving out counts reported by others; Epoch often adds counts for recent models late, and we have not checked the blanks against the labs' own pages.
Parameter counts reported by others are left out of the parameter figures. Record updated 2026-10-19.