Epoch AI Benchmarking Hub · 2024-09-19 to 2026-02-17
8
+47.5
64 d
8.8
Lab race
Which lab held the capability frontier at each point in time, and how often the lead changed hands.
Each line is the best consensus score among a lab's models released on or before that date. The leader on a date is the lab with the highest line. Scores are today's, placed at release dates, not past snapshots. 38 models without a release date are left out.
Leads today
Google
Gemini 4 Carbon, since Nov 10, 2026
Lead over second
+4.8
Intervals overlap
Lead changes
17
4.6 a year, 16 within noise
Longest reign
465 d
OpenAI
Lead changeHover a point for the model that set the step. Gray lines are other labs.
Lead changes (17)
2024-06-21: OpenAI to Anthropic with Claude 3.5 Sonnet (June '24), +0.1 (within noise)
2024-09-12: Anthropic to OpenAI with o1-preview, +6.1 (within noise)
2025-01-20: OpenAI to DeepSeek with DeepSeek R1 (Jan '25), +0.3 (within noise)
2025-01-31: DeepSeek to OpenAI with o3-mini, +0.1 (within noise)
2025-03-25: OpenAI to DeepSeek with DeepSeek V3 0324, +0.1 (within noise)
2025-04-16: DeepSeek to OpenAI with o3, +8.7 (within noise)
2025-08-05: OpenAI to Anthropic with Opus 4.1, +1.7 (within noise)
2025-08-07: Anthropic to OpenAI with GPT-5, +3.0 (within noise)
2025-09-29: OpenAI to Anthropic with Sonnet 4.5, +0.7 (within noise)
2025-11-13: Anthropic to OpenAI with GPT-5.1, +0.4 (within noise)
2025-11-18: OpenAI to Google with Gemini 3 Pro Preview, +10.8 (within noise)
2026-02-05: Google to Anthropic with Opus 4.6, +14.9
2026-03-05: Anthropic to OpenAI with GPT-5.4 Pro, +1.0 (within noise)
2026-06-09: OpenAI to Anthropic with Fable 5, +0.5 (within noise)
2026-10-13: Anthropic to Google with Gemini 4 Argon, +0.4 (within noise)
2026-11-09: Google to Anthropic with Fable 5.2, +2.5 (within noise)
2026-11-10: Anthropic to Google with Gemini 4 Carbon, +4.8 (within noise)
Days in the lead
OpenAI913 d (68%)
took the lead 7 times
Anthropic285 d (21%)
took the lead 6 times
Google113 d (8%)
took the lead 3 times
DeepSeek33 d (2%)
took the lead 2 times
18 other labs never led.
Method, limits, and excluded inputs (34 boards, 3509 rows read)
Method rules
Score: the latent capability (0-100) of each model from the consensus fit, with its 95% bootstrap interval.
Date: the model's registry release date. Models without a valid release date are left out and counted.
Per lab: a step curve of the best score among its models released on or before each date; a step is drawn only when a release beats the lab's earlier best.
Leader: on each release date, the lab with the highest best score. A tie keeps the current leader; a tie on the first date goes to the lab with the lower id.
Lead change: a date on which a different lab becomes the leader. The margin is the new leader's best score minus the old leader's best score on that date.
Days in the lead: whole days from the date a lab took the lead to the date the next lab took it, or to the reference date for the current leader.
Gap to the leader: the leader's best score minus the lab's best score today.
Known limits
Scores come from today's data, not from past snapshots: each model is placed at its release date with the score it has now, so the race shows what we know today about the past, not who looked ahead at the time.
A lab's score changes with the set of models and boards in the fit, so a new release can shift every score; the lead is relative to today's fit.
A margin smaller than the overlap of two 95% intervals is within noise; lead changes with overlapping intervals are flagged `withinNoise`.
Only models with enough boards for a consensus score are counted, so an early or narrow-coverage model may be missing from its lab's curve.
Release dates are the registry's dates; a model that was announced earlier than it was released, or tested before release, is placed at its release date.
Exclusions
38 models excluded: model has no valid release date
Price of fixed capability
For each board with a cost, three score tiers (the median, top quarter and top tenth of the models on that board). Each line is the cheapest cost, among models released so far, of any run at or above the tier. A halving time comes from a log-linear fit, and only for a curve with at least 4 steps over 180 days; a board needs 8 dated, priced models. Costs are as each source reports them, so compare the pace, not the level.
Median halving time
3 months
2 fitted tiers on 13 boards with a cost
Only board with a fit
ARC-AGI-2 (Epoch AI Benchmarking Hub)
The halving time rests on this one board. Do not read it as a market-wide pace.
Median scorescore at least 5.9, 30 modelshalves every 4.4 months, 85% cheaper per year
Top quarterscore at least 53.22, 15 modelshalves every 1.6 months, over 99% cheaper per year
Top tenthscore at least 75.93, 6 modelsno fit (2 steps)
ARC-AGI-2: score is Score, cost is Cost per task (unknown cost), 60 dated models released 2024-07-18 to 2026-07-16. Hover or focus a point to see the model that set the step.
Method, limits, and excluded inputs (13 boards, 693 rows read)
Method rules
A board is used when it is the default slice of its source, its rank metric is a quality score, and it has a cost metric apart from the rank metric.
Rows count when they are visible and have a score, a positive cost, a registry release date, and the cost kind that most rows of the board share (billed, list, floor or unknown are never mixed).
A board needs at least 8 such models. Its 3 tiers are the median, the 75th and the 90th percentile of the best score of each model.
For a tier, every row at or above the tier score is a candidate. Sorted by release date, a row makes a new step when its cost is strictly below every earlier candidate; on one date the cheapest row wins.
A log-linear fit (least squares of ln cost on years) runs only when a curve has at least 4 steps spanning at least 180 days. Cost change per year is exp(slope) - 1; halving time is 12 * ln 2 / -slope months.
The summary is the median halving time over every fitted tier, and the fastest and slowest boards by the median halving time of their fitted tiers.
Known limits
Cost is a per-task, per-test or per-run figure as each source reports it, so boards are not comparable on level, only on pace.
Release dates come from the model registry only; a model without one is left out (an observation date would only mark when a source saw the model).
A step curve falls by construction, so the fit measures how fast it falls, not whether it falls. Few steps make the halving time unstable.
Tiers are cut from the models a board happens to hold, so the same tier name is a different score on each board.
Boards from one source (for example the Vals family) share models and prices, so their halving times are not independent.
Exclusions
1 boards excluded: board has fewer than 8 dated, priced models
36 boards excluded: board has no cost metric apart from its rank metric
94 boards excluded: not the default slice of its source and version
1 boards excluded: older version of a benchmark the source reports again
2 boards excluded: rank metric is not a quality score (price, speed, latency, or token count)
37 curves excluded: curve has fewer than 4 steps over 180 days (no halving time)
1 rows excluded: cost is zero or negative (free rows cannot sit on a log axis)
36 rows excluded: cost kind differs from the main cost kind of the board
75 rows excluded: hidden row (for example superseded)
48 rows excluded: model has no registry release date
163 rows excluded: row has no cost (not tested for cost)
Method, limits, and excluded inputs (34 boards, 3509 rows read)
Method rules
Boards: default quality boards with at least 6 dated models (best visible row per model).
Each model uses its registry releaseDate when present, falling back to row observedAt when no release date exists.
At each date, the SOTA frontier records every release that sets a strict new best score overall, within open weights, and within proprietary models.
Open-weight time lag measures days between when proprietary models first reached today's open-weight peak and when open weights reached it.
Known limits
Models without a known release date or observation date are excluded from timeline curves and listed in exclusions.
Weight availability for older legacy models without an explicit flag defaults to unknown and is omitted from the open-vs-proprietary split.
Exclusions
16 boards excluded: board has fewer than 6 dated models
94 boards excluded: not the default slice of its source and version
1 boards excluded: older version of a benchmark the source reports again
2 boards excluded: rank metric is not a quality score
18 models excluded: model has no releaseDate or observedAt
75 rows excluded: hidden row (for example superseded)