Every AI model on one ruler.
352 ranked models stand on it today. The top climbs about 1.2 SCI a month.
Source: Synopticon Capability Index (SCI), an item-response fit over 908 benchmarks using Epoch AI's original 2025 ECI method. 251 of 352 ranked models are placed by release date (the leaderboard's dates: first public or partner release; the chart starts in 2023). Pace: least-squares fit to the 15 records set in the last 24 months, on the leaderboard's release dates. Claude 3.5 Sonnet = 130, GPT-5 (high) = 150. Data as of 11 Nov 2026. The opening cloud is every score we hold, one dot each: 62,835 of 64,602 sit on this ruler (the rest come from models off it or not yet fitted). Each dot sits at its model's SCI and the share of its benchmark solved above chance, placed at random inside a 0.5-SCI by 2-point cell.
Evidence note
The board ranks each model by its best configuration and prints a badge when evidence is thin or lab-reported only; this page shows those badges as they are. We call a lead or a change of lab by less than 3 points too close to call. The daily fit publishes point estimates, so no per-model interval is shown here. Every model, with its configurations.
How fast
Three rates and a gap, each with its source and its date. Two are refreshed every day with the ruler; METR's horizons and the price snapshot carry the date of their last release.
Open weights run 5.1 months behind
Best SCI to date, on third-party results only (lab-reported scores left out), released models. The best open-weight model, Kimi K3.1 Pro, scores what Claude Fable 5 reached in Jun 2026. Shaded is capability only closed models had; today's gap is 6.4 points.
Source: Synopticon Capability Index, third-party-only fit (lab-reported scores left out), 88 open- and 95 closed-weight models; open weights per Epoch AI where it has a flag. Data as of 11 Nov 2026.
The length of task AI can finish doubles every 4.2 months
Skilled-human time for software tasks a model completes half the time; all 23 models on METR's Time Horizon 1.1, log scale. The doubling time is METR's own fit from 2023 on (METR's interval: 3.4–5.2 months).
Source: METR Time Horizon 1.1, results through models released Apr 2026, fetched 16 Oct 2026; task examples from METR papers.
GPT-4o-level capability got 29 times cheaper in a year
Cheapest full run of Artificial Analysis's test suite at GPT-4o's score or better, at Jun 2026 list prices, log scale. The fitted decline is 85% a year (90% CI 65–95%), but the cheapest price has been flat since Aug 2025.
Source: Artificial Analysis eval costs, 7 Jun 2026 snapshot, the last before AA revised its index in June 2026 (later snapshots use a new scale and are not mixed in).
Labs' own scores run from −0.1 to +2.8 points higher than independent ones
Median of a lab's self-reported score minus a third-party score on the same model and benchmark, labs with 20 or more like-for-like pairs (593 pairs). Points on a 0–100 benchmark scale.
Source: Synopticon score store, like-for-like pairs, data as of 11 Nov 2026. Every lab: /inflation/.
Every lab, and the largest gaps →Every time the record fell
17 new records since Gemini 1.5 Pro. The lead changed labs 11 times; 6 of those changes were by less than 3 points, too close to call.
A heavy rule marks each change of lab; a dashed one, a change by less than 3 points. The arrow marks the record that still stands. Badges are the leaderboard's own. Source: Synopticon Capability Index, best configuration per model, leaderboard release dates, data as of 11 Nov 2026.
Research
Longer pieces built on the same data.
Many viewpoints, one record.
The AI race, told through evidence. Collected at the source, every day.
Most of the AI story is told through announcements. The evidence underneath (benchmark results, prices, token volumes, developer adoption) lives in dozens of places and changes weekly.
Synopticon collects it every day and keeps a continuous record. The Synopticon Capability Index fits every benchmark result we can match into one capability score per model, and weighs an independent measurement more heavily than a lab's own claim. Where the sources disagree, we publish the gap.
The method is Epoch AI's original 2025 one. We fit the least-squares model behind their Epoch Capabilities Index (ECI), from A Rosetta Stone for AI Benchmarks, and anchor it to the same two models, so an SCI reads in ECI units. Epoch has since moved ECI to a Bayesian fit (November 2026). SCI is our own fit over many more benchmarks: not Epoch's published index, and not affiliated with Epoch AI.
Get the weekly research
New findings from the data, once a week.