AI labs · October 2026
Everyone else is three months behind Claude and GPT
Since September 2024, OpenAI or Anthropic has had the top-scoring AI model on 96% of days. The challenger keeps changing. The gap mostly doesn’t, though in October Meta got within weeks.
On 12 Oct, Meta released Muse Wonder Full. It scores 55.8 on the Artificial Analysis Intelligence Index, a blend of hard tests in reasoning, knowledge, coding and agent work. That’s 1.8 points behind the best, Claude Opus 5.5 at 57.6, which Anthropic released 20 days earlier.
That’s as close as anyone outside OpenAI and Anthropic has come in almost a year. Usually they are about three months behind, and have been since reasoning models arrived in 2024. Press play and watch the top of the chart.
Anthropic, 57.6
The best score among models each lab had released by that date, on the Artificial Analysis Intelligence Index (current version, highest effort setting). The axis rescales as scores rise; the top 10 labs are shown.
Below the top two, the order churns. Eight labs have had the best model outside OpenAI and Anthropic since Sep 2024: Meta, Alibaba, Google, DeepSeek, Perplexity, xAI, Z.ai and Kimi. At the top, it’s Claude or GPT on all but 32 of 778 days. The exceptions: Grok 4 from xAI, for 26 days in summer 2025, and Gemini 3 Pro Preview from Google, for 6 days that November, before Claude Opus 4.5 passed it.
Artificial Analysis Intelligence Index (current version) of the best model released by each date. Blue shading is Claude and GPT’s lead; orange, the two spells when someone else led. Hover or tap for the models.
Three months, give or take
A clearer way to see the gap is in time. On any given day, the best model from outside the two labs matches what Claude or GPT had some number of days earlier. In 2025 the median was 83 days; this year it’s 81. The number swings as each side ships, from a few weeks to most of a year, but it doesn’t trend toward zero.
Days since OpenAI or Anthropic first released a model scoring at least as high as the best model from any other lab, weekly. Below zero: another lab led, and the number is days until Claude or GPT passed it.
It’s not one leaderboard
One index can flatter whoever it was built around, so we checked Epoch AI, a research group that runs its own tests on new models and collects results from others. It tracks 35 benchmarks with at least 20 models from 5 labs since September 2024. On 11 of them, nobody but OpenAI or Anthropic has ever had the top score. On 23, they had it at least 85% of the time.
Google had the top score for a while on 11 of the other 12. The five where Claude and GPT did worst, GeoBench, Balrog, SimpleBench, GPQA Diamond and Lech Mazur Writing, test placing a photo on a map, playing video games, trick questions, graduate-level science and short-story writing: less of the long, careful work of coding and math.
Share of days, from September 2024 (or the benchmark’s first model) to its newest model, on which the highest-scoring model came from OpenAI or Anthropic. Data: Epoch AI (CC BY 4.0), fetched 21 Oct 2026. Hover or tap a row for who else led.
The one coders care about
SWE-bench Verified asks a model to fix real bugs from open-source projects on GitHub, and is the test most AI coding tools quote. Epoch ran it on 32 models from 7 labs, all with the same setup, until June. Claude or GPT held the top score the whole time. The closest call: Gemini 3 Pro, at 72.9%, came 0.6 points short of GPT-5.
Best score of any model released by each date, in Epoch AI’s own runs (20 Nov 2024 to 16 Jun 2026, when Epoch’s runs end). Epoch never ran xAI’s Grok on it. Hover or tap for the models.
Where everyone else wins
Price. MiMo-V2.6-Pro, from Xiaomi, scores 46.3 and costs 54 cents per million tokens. No Claude or GPT model that scores as high costs less than $4: about 7 times more. Three months behind the best is still better than any model was a year ago, and for most work it’s plenty.
The other labs also ship open weights, which OpenAI and Anthropic mostly don’t, and Google leads on several knowledge tests. If you want the very best model on the day, though, the last two years say it will come from one of two companies.
Method
Intelligence scores are the Artificial Analysis Intelligence Index (current version), taking each model at its highest effort setting, by release date, as of 18 Oct 2026. A model counts from its release date. We count from September 2024 because before that every model scores between 5 and 9 on the current index, too close to rank. The race starts at GPT-4 on 14 March 2023 and shows each lab’s best model on each date. “Days behind” is the number of days since OpenAI or Anthropic first released a model scoring at least as high as the best model from anyone else.
Benchmarks are from Epoch AI’s Benchmarking Hub (CC BY 4.0), fetched 21 Oct 2026: every benchmark in its metadata with a score column, not superseded by a newer version, with at least 20 models released since September 2024 from at least 5 labs. Some are Epoch’s own runs; others collect results from the benchmark’s authors. A model holds the top score from its release date until a later model scores higher. Each benchmark is counted to the release date of the newest model on it, so one Epoch stopped running isn’t counted after that. Models are credited to the first organisation Epoch lists, except that a model co-credited to OpenAI or Anthropic counts for them.
Two caveats. OpenAI funded FrontierMath and has seen many of its problems, and Epoch’s current version of it wasn’t run on Gemini 3 Pro; Google models led the earlier versions for about a month in total. SWE-bench Verified ends in June 2026, and Grok was never run on it. The data and code are in our repository (data/epoch-history, scripts/epoch-leads.ts).