OpenAI First on 20 AI Benchmarks, Anthropic Leads 2026 at 11

OpenAI’s models were the first to reach 85% on 20 widely used AI benchmarks, more than any other lab, and Anthropic leads for 2026 with 11, according to SatBench, a tracker that Ethan Mollick posted on X on 14 November.

Mollick wrote that “Of 173 benchmarks cited by the Labs since 2023, 63% are now saturated.” His chart’s footnote adds that counting only independent runs, 24% are solved. By our arithmetic, that leaves about 67 of the 109 solved benchmarks without an independent run at 85%.

Business Pill · Benchmarks explained

An evergreen Business Engineer explainer on benchmarks and why a test score is not your own result. It teaches the general idea only and does not discuss SatBench.

The key insight: By our arithmetic, roughly 6 in 10 of the benchmarks SatBench counts as solved reach 85% only when non-independent scores, such as labs’ own launch-post results, are included: about 67 of 109. The leaderboard of who got there first rests on the same first recorded scores.

What SatBench Counts

SatBench covers the benchmarks named in 10 or more AI lab launch posts. A benchmark counts as solved once the best AI score reaches 85%, and the best score can come from any source, including labs’ own launch posts.

On that basis, 109 of the 173 benchmarks were solved as of 14 November, the data date on the charts, which cite Benchmark Almanac and Epoch AI’s Benchmarking Hub. The post’s text puts it as: “Saturation is happening faster each year.”

SatBench’s 173 lab-cited AI benchmarks as of 14 November 2026, split into those solved in independent ru
SatBench’s 173 lab-cited AI benchmarks as of 14 November 2026, split into those solved in independent runs (about 42, the footnote’s 24%), those solved only with other scores (about 67) and those unsolved (64). The split is our arithmetic from Ethan Mollick’s charts.

What Was Posted on X

Ethan Mollick posted on 14 November at 23:46 UTC that he had built SatBench to measure how quickly major AI benchmarks become saturated, and that 63% of 173 benchmarks cited by the labs since 2023 are now saturated. The four charts attached to the post carry the leaderboard, the 24% independent-run footnote and the unsolved list this piece uses; the split of about 67 benchmarks is our own arithmetic on those charts.

Who Got There First

The leaderboard credits a lab when its model was the first recorded to reach 85% on a benchmark several frontier labs report. Of 71 such benchmarks, OpenAI has 20, Google 14 and Anthropic 14, then Meta, ByteDance and Qwen with 4 each and all other labs 11.

By our arithmetic, OpenAI holds 28.2% of the 71, and OpenAI, Google and Anthropic together hold 48, or 67.6%. The chart text says Anthropic leads in 2026 with 11, which is 11 of its 14 firsts, or 79%, by our arithmetic.

The chart names SWE-bench Verified and SWE-Bench Pro as examples for Anthropic, MMLU-Pro and MMLU for OpenAI, and GPQA Diamond and BrowseComp for Google.

How Much Rests on Independent Runs

The footnote’s 24% applies to all 173 benchmarks, which is about 42, by our arithmetic. Set against the 109 solved, about 67, roughly 6 in 10, reach 85% only when non-independent scores such as labs’ own launch-post results are counted.

Because 24% is rounded, the independent count could be 41 or 42, so the gap is 67 or 68. How to read a published test score against your own results is in our benchmarks explainer.

What Is Still Unsolved

The charts list 64 benchmarks as unsolved. Humanity’s Last Exam is the top holdout at 74%, and MMMU-Pro, SciCode and IFBench are also unsolved.

The share of new benchmarks solved within a year of release rose from 17% for 2023 tests to 30% for 2024 and 43% for 2025, 2.5 times in two years by our arithmetic. For 2026 tests the chart gives 32% within six months.

The Structural Read

As we read the leaderboard, Anthropic leads 2026 with 11 of its 14 firsts, while OpenAI’s and Google’s bars are mostly in the before-2026 segment of the chart’s split.

As we read the footnote, the gap between 63% and 24% is a measure of how much of the progress record rests on non-independent scores. The tests SatBench counts are the ones named in 10 or more lab launch posts.

Ethan Mollick, on X, 14 November 2026

“AI is getting better faster than our measurements are”

Three Implications

FOR BUYERS OF AI MODELS SatBench’s footnote gives 24% solved when only independent runs count, against 63% when labs’ own launch-post scores are included.

FOR THE LABS OpenAI was first to 85% on 20 of the 71 leaderboard benchmarks, Google and Anthropic on 14 each; Anthropic leads 2026 with 11, according to SatBench.

RELATED How to read a published score against your own results is in our benchmarks explainer.

The Business Engineer Lens

This story maps onto the Business Engineer framework The Product Overhang Doctrine.

The framework argues: “Most product organizations are built to ship for the curve they can see today.”

As we read it, SatBench measures the other side of that curve: benchmarks that labs cite are being passed faster each year, which widens the distance between tested capability and shipped products. The framework predates SatBench and does not discuss it.

What Is Not Established

SatBench is one researcher’s compilation, which Mollick describes as thrown together, from third-party trackers. The 85% line is its own definition, and the leaderboard rests on first recorded scores, including results labs report about their own models.

Its chart on when the full set is saturated, around 2028 or 2030, is labelled an extrapolation and “Not a forecast”. We did not contact OpenAI, Anthropic, Google or Mollick for this piece.

This Week In The AI Supercycle

AI Lost Its Central Banker, the Lab Oligopoly, and the Future of Open Source

This week’s issue: the Nvidia leadership change, read as the AI build-out losing its central banker: who now finances the GPU landlords, backstops the data centres and sponsors open models, and why two labs are becoming the bank.

Read the full issue →

Business Engineer Framework

The Product Overhang Doctrine

The Business Engineer analysis of the gap between what frontier models can do and what products ship.

Read the premium analysis →Get 50% off →

The Bottom Line

SatBench counts 109 of 173 lab-cited AI benchmarks as solved, with OpenAI first on 20 and Anthropic leading 2026 with 11. Counting only independent runs, its footnote gives 24%, which leaves about 67 solved benchmarks without an independent run at 85%, by our arithmetic.

Anthropic’s share offering is mapped in our Anthropic IPO map.

95,000+ executives read Business Engineer for the AI strategy frameworks cited by ChatGPT, Claude, and Perplexity.

A note on sourcing. This piece rests on Ethan Mollick’s post on X of 14 November 2026 and the four SatBench charts attached to it, which cite Benchmark Almanac, including Epoch AI’s Benchmarking Hub, with data as of 14 November. We read the post in full and transcribed the charts’ text; we did not rerun the underlying data. Shares, the split of solved benchmarks and growth multiples are our arithmetic. OpenAI, Anthropic, Google and Mollick were not contacted for this piece.

Sources: Ethan Mollick on X, SatBench, 14 November 2026 · Benchmark Almanac (data cited by SatBench)

Scroll to Top
FourWeekMBA