Is AI progress accelerating?
Better models keep arriving. But is the pace picking up? Follow the strongest released models, then compare how much ground they gain over equal periods of time.
What the data says
Over the year: faster progress. Recently: slower progress.
Last 12 months
17.9ECI points / year
Previous period: 15.8 +13.7%
See comparison dates
Latest period: November 8, 2025 – November 8, 2026
Previous period: November 8, 2024 – November 8, 2025
Last 6 months
12.7ECI points / year
Previous period: 23.2 -45.4%
See comparison dates
Latest period: May 8, 2026 – November 8, 2026
Previous period: November 8, 2025 – May 8, 2026
How capable are the best models?
Highest score to date169.1 ECIGPT-6 Astra
Each dot is a released model. The line follows the highest score reached so far. A step upward marks a new frontier.
Capability · ECI score
229 released models · 22 frontier milestones
How fast are they improving?
A rising line means more capability gained per year. A falling line means slower improvement. Both windows use the same vertical scale.
Pace · ECI points gained per year
Pace · ECI points gained per year
Blank at the beginning means there is not yet a full comparison window. The line can fall while model capability continues to rise.
The models that moved the frontier
A new milestone appears only when a release beats the previous highest score. Dates follow Epoch AI’s dataset.
Explore the frontier milestones
| Model | Release date | ECI | 90% credible interval |
|---|---|---|---|
| GPT-6 AstraOpenAI | 169.1 | 166.2–172.3 | |
| Claude Fable 5.1Anthropic | 166.9 | 164.3–169.8 | |
| Claude Opus 5Anthropic | 164.4 | 162.0–167.2 | |
| Claude Fable 5Anthropic | 164.2 | 161.8–166.9 | |
| GPT-5.5 ProOpenAI | 162.7 | 160.0–165.6 | |
| GPT-5.4 ProOpenAI | 160.2 | 157.7–162.9 | |
| GPT-5.2 ProOpenAI | 157.2 | 155.0–159.7 | |
| Gemini 3 ProGoogle DeepMind | 152.9 | 151.3–154.6 | |
| GPT-5 ProOpenAI | 151.2 | 149.3–153.1 | |
| GPT-5OpenAI | 150.0 | Not provided | |
| o3-proOpenAI | 147.4 | 145.6–149.3 | |
| o3OpenAI | 147.0 | 145.6–148.5 | |
| Gemini 2.5 Pro (Mar 2025)Google DeepMind | 143.9 | 141.9–145.8 | |
| o1OpenAI | 141.2 | 139.5–142.8 | |
| o1-miniOpenAI | 135.4 | 133.1–137.5 | |
| Claude 3.5 SonnetAnthropic | 130.0 | Not provided | |
| GPT-4o (May 2024)OpenAI | 128.2 | 125.5–130.8 | |
| GPT-4 Turbo (Apr 2024)OpenAI | 125.8 | 122.7–128.7 | |
| Claude 3 OpusAnthropic | 125.6 | 122.6–128.4 | |
| GPT-4 Turbo (Nov 2023)OpenAI | 124.8 | 121.6–127.8 | |
| GPT-4 (Mar 2023)OpenAI | 121.9 | 118.0–125.2 | |
| LLaMA-65BMeta AI | 102.0 | 93.5–109.4 |
How to read this
One consistent capability scale
The Epoch Capabilities Index (ECI) combines benchmark results into a shared scale. A higher score means stronger measured performance; it is not a percentage of intelligence.
We include public API, hosted, and open-weight releases. Unreleased, limited-access, unknown-accessibility, and future-dated models are excluded.
Compare equal periods
Pace is the highest-score gain over the preceding 6 or 12 calendar months, converted to points per year using actual elapsed days. Each rate is compared with the immediately preceding equivalent period.
More than 10% faster or slower earns that label; changes within ±10% count as a similar pace. This is a descriptive convention, not a statistical significance test. Percentages compare pace, not intelligence.
An estimate, frozen in time
The model tooltips and table show Epoch’s 90% credible intervals where provided. These describe score uncertainty, not confidence that progress is accelerating. Benchmark coverage and release timing also affect the picture.
All calculations end at the snapshot date. Updates are manual and monthly. Each refresh replaces the full dataset because historical estimates can change. We make no forecasts.
Source: Epoch AI. Trend calculations: Generative Networks. Data licensed under CC BY 4.0.