A post by Alex Michael at 20:55 UTC on 7 November charting GPT-4o through GPT-6 Astra on the same independent benchmarks, quoting his earlier post that does the same for Claude Opus 3 through Opus 5.5. He says all results come from Epoch AI and the benchmark maintainers and none are self-reported by the labs. The compilation is his; we checked each figure used here against Epoch’s data files.
On Epoch AI’s Capabilities Index, Anthropic’s Claude went from 125.6 with Claude 3 Opus in February 2024 to 168.4 with Opus 5.5 in September 2026, a gain of 42.8 points by our arithmetic. OpenAI’s GPT went from 128.9 with its best 2024 GPT-4o snapshot to 169.1 with GPT-6 Astra, a gain of 40.2 points. The two newest models are now 0.7 points apart.
The figures come from Epoch AI’s public benchmarking data, which we downloaded on November 8. Alex Michael compiled them into two tables on X, one for each lab, and says “All results come from @EpochAIResearch and the benchmark maintainers.” The footnotes on his charts say each model is shown “at the best reasoning setting tested.”
Business Pill · The Model and the Harness
An evergreen Business Engineer explainer on how the harness around a model shapes what it delivers. It teaches the general idea only and does not discuss these benchmark scores.
The key insight: On Epoch AI’s Capabilities Index, Claude gained 42.8 points from Claude 3 Opus to Opus 5.5 and GPT gained 40.2 from GPT-4o to GPT-6 Astra, by our arithmetic, leaving the two newest models 0.7 points apart with overlapping 90% intervals.
What the Two Tables Compare
Alex Michael’s GPT table runs GPT-4o, o3, GPT-5, GPT-5.5, GPT-6.1 Sol and GPT-6 Astra; his Claude table runs Opus 3, 4, 4.5, 5 and 5.5. His footnotes say the scores are “independent evaluations compiled by Epoch AI” and not supplied by either company. In Epoch’s files, the ARC-AGI rows are credited to the ARC Prize, via Epoch (Epoch AI).
We checked every figure used here against Epoch’s downloadable data. The GPT-4o index figure of 128.9 is the November 2024 snapshot, the best of OpenAI’s three 2024 GPT-4o versions; the May 2024 version scores 128.2.
Score of each lab’s 2024 flagship (GPT-4o, Claude 3 Opus) and its newest model (GPT-6 Astra, Claude Opus 5.5) on four tests Epoch AI scores for all four models, each at its best setting tested. Data: Epoch AI benchmarking data, retrieved 8 Nov 2026; compiled by Alex Michael.
Where Each Lab Gained More
We compared the two labs only on tests where Epoch has a score for both the 2024 flagship and the newest model. On four of those, by our arithmetic:
Competition math (OTIS Mock AIME 2024-2025): Claude rose from 4.7% to 100.0%, up 95.3 points; GPT from 6.4% to 100.0%, up 93.6. Factual recall (SimpleQA Verified): Claude from 12.6% to 72.2%, up 59.6; GPT from 26.0% to 75.6%, up 49.6. Chess puzzles: GPT from 9.0% to 74.0%, up 65.0; Claude from 4.0% to 67.0%, up 63.0. Graduate-level science (GPQA Diamond): GPT from 49.2% to 95.8%, up 46.6; Claude from 47.2% to 90.6%, up 43.4.
Counting the index as well, Claude gained more on three of the five comparisons and GPT on two. Claude started lower on every one of the five.
Where They Stand Today
On seven tests where Epoch scores both newest models, GPT-6 Astra is ahead on five and the two tie on two. Astra leads on the index (169.1 to 168.4), graduate-level science (95.8% to 90.6%), factual recall (75.6% to 72.2%), chess puzzles (74.0% to 67.0%) and ARC-AGI-2 (95.0% to 93.3%). Both score 100.0% on competition math and 98.5% on ARC-AGI-1.
Claude’s best graduate-science score in the table belongs to the older Opus 5, at 93.9%, not Opus 5.5. Epoch’s index gives each newest model a 90% interval: 166.16 to 172.28 for GPT-6 Astra and 165.54 to 171.51 for Opus 5.5, so the two ranges overlap. We reported on November 5 that Epoch put a 72% chance on Astra ranking above Opus 5.5.
Points per Year
Measured from each lab’s first 2024 flagship in Epoch’s data, by our arithmetic: GPT-4o’s May 13, 2024 version (128.2) to GPT-6 Astra on September 3, 2026 (169.1) is about 17.7 index points a year over 2.31 years. Claude 3 Opus on February 29, 2024 (125.6) to Opus 5.5 on September 22, 2026 (168.4) is about 16.7 points a year over 2.56 years.
The two tables do not use the same full set of tests. The GPT table includes research-level math (FrontierMath) and research-level physics (CritPt); the Claude table includes SimpleBench, WeirdML and METR’s autonomous task length, with no GPT row shown for those. Several tests are near their ceiling: both newest models score 100.0% on competition math and 98.5% on ARC-AGI-1, so those rows can no longer separate them.
By our arithmetic, Claude gained 42.8 Epoch index points from Claude 3 Opus to Opus 5.5 and GPT gained 40.2 from its best 2024 GPT-4o snapshot to GPT-6 Astra; the newest models are 0.7 points apart.
The Structural Read
As we read the two tables, the larger story is the size of the jump on both sides rather than the order: on competition math both labs went from single digits to 100.0%, and on chess puzzles from under 10% to about 70%.
As we read Epoch’s intervals, a 0.7-point gap between 169.1 and 168.4 does not settle which model is ahead, and Epoch itself put the probability at 72% for Astra, as we reported on November 5.
As we read the benchmark lists, each table also carries tests the other lab is not scored on, so a reader comparing only the headline rows sees a narrower slice than either table shows.
Alex Michael, on X
“All results come from @EpochAIResearch and the benchmark maintainers.”
Three Implications
FOR BUYERS CHOOSING A MODEL On the seven tests Epoch scores for both newest models, the largest gap is 7.0 points on chess puzzles, by our arithmetic, and two tests are tied, so the tables alone do not separate them on most rows.
FOR READERS OF LEADERBOARDS The figures are each model at its best setting tested, which Alex Michael’s footnotes state, so a default setting in a product can score lower.
FOR TRACKING THE RACE By our arithmetic, GPT added about 17.7 index points a year and Claude about 16.7 from their first 2024 flagships in Epoch’s data.
The essay argues that “Model weights are increasingly fungible” and that what separates the labs now is “the harness — the agentic scaffold around the model”.
As we read it, Epoch’s 0.7-point gap fits that description of models clustering at the top of the same tests. The essay was published before GPT-6 Astra and Opus 5.5 and does not discuss these scores.
What Is Not Established
These are scores on benchmarks, not measures of how the models perform in a given product or company. The tables show each model at its best setting tested, which can mean more compute per answer. The tables do not include price; we have covered cost separately, including a cheap model handing hard cases to Opus. Opus 5.5 is not scored on SimpleBench, WeirdML or METR’s task-length test in the Claude table, and the 0.7-point gap sits inside both models’ 90% intervals.
This Week In The AI Supercycle
The Great Convergence of Capital, Power, and Geopolitics
This week’s issue: for two years the AI build-out was priced as a supply story, and last week as a credit story. This week it was priced as a permission story, and the answer arrived as a cost, not a ban.
Over the same independent tests, Claude gained 42.8 index points since early 2024 and GPT gained 40.2, leaving GPT-6 Astra 0.7 points ahead of Opus 5.5 on Epoch’s index. GPT-6 Astra leads on five of seven shared tests and ties on two.
A note on sourcing. Scores are from Epoch AI’s public benchmarking data, downloaded on 8 November 2026, as compiled by Alex Michael on X; we checked each figure against Epoch’s files. We did not rerun any evaluation. Gains, gaps and points per year are our arithmetic.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
