Epoch AI says GPT-6 Astra has “a 72% probability of being more capable than Opus 5.5, but a 99% probability of being more capable than Fable 5.1.” The research group posted the figures on X on 4 November, after moving its Epoch Capabilities Index (ECI) to a Bayesian method.
On OpenAI’s and Anthropic’s pricing pages, GPT-6 Astra lists at $10 per million input tokens and $50 per million output tokens, and Claude Opus 5.5 at $4 and $20. That makes Astra 2.5 times Opus 5.5’s price on both, by our arithmetic.
Business Pill · Percentiles
A short Business Engineer explainer on percentiles, the idea behind the 90% intervals in Epoch’s index. It teaches the general idea only and says nothing about Epoch or these models.
The key insight: Epoch AI gives GPT-6 Astra a 72% probability of being more capable than Claude Opus 5.5 and 54% against GPT-6.1 Sol; on list prices, Astra costs 2.5 times Opus 5.5 and five times Sol, by our arithmetic.
Epoch writes that it has “updated the Epoch Capabilities Index (ECI) to use a Bayesian methodology.” It says this “improves uncertainty estimates and lets us compare models more accurately, even when their individual ECI uncertainty intervals overlap.”
It adds that “frontier model scores are largely unchanged: models stronger than GPT-4 move by less than 1 ECI point on average.” The new part is the head-to-head comparison: with “a joint distribution over all models’ ECIs, we can now compare models directly.”
Epoch Capabilities Index point estimates and 90% credible intervals for six frontier models, from Epoch AI’s scores file updated 5 November 2026, with each model’s list output price per million tokens from OpenAI’s and Anthropic’s pricing pages. Gemini 4 Argon is omitted because our source has no price for it.
What Was Posted on X
With better uncertainty estimates and a joint distribution over all modelsโ ECIs, we can now compare models directly.
For example, we can now say Astra has a 72% probability of being more capable than Opus 5.5, but a 99% probability of being more capable than Fable 5.1.
Epoch AI posted this on 4 November with a chart. It says that with a joint distribution over all models’ ECIs it can compare models directly, and that Astra has a 72% probability of being more capable than Opus 5.5 but a 99% probability of being more capable than Fable 5.1.
Astra 6.2 has a 72% probability of being more capable than opus 5.5 and a 99% probability of being more capable than Fable 5.1
whoa i bet Astra 6.2 will score even higher than fable 5.5 and opus 5.6 and i guess that it ends up at 173 ECI
Haider quoted Epoch’s post on 5 November, describing the model as Astra 6.2, and guessed that it ends up at 173 ECI. That figure is Haider’s own guess; Epoch’s chart and scores name GPT-6 Astra.
The Head-to-Head Numbers
Epoch’s chart is titled: “GPT-6 Astra is very likely ahead of three of the next five models, but its lead over the closest two is uncertain.” It gives the probability that GPT-6 Astra is more capable as 54% against GPT-6.1 Sol and 72% against Claude Opus 5.5.
Against Claude Fable 5.1 and Claude Sonnet 5.5 the figure is 99%, and against Gemini 4 Argon 100%. The chart notes that “gaps are computed draw by draw, which accounts for the models’ ECIs moving together,” with data as of 26 October 2026.
Epoch’s downloadable scores, labelled as updated on 5 November, put GPT-6 Astra at 169.07, with a 90% credible interval of 166.16 to 172.28. GPT-6.1 Sol is at 168.94, Claude Opus 5.5 at 168.36, GPT-6.1 Terra at 167.82 and Claude Fable 5.1 at 166.90.
What the Price Gap Buys
On Epoch’s point estimates, Astra sits 0.71 ECI points above Opus 5.5 and 0.13 above GPT-6.1 Sol, by our arithmetic. Sol lists at $2 and $10, one fifth of Astra’s price, and Epoch puts the chance that Astra is more capable than Sol at 54%.
Claude Fable 5.1 lists at the same $10 and $50 as Astra. Epoch puts Fable 2.17 points lower on its point estimates, by our arithmetic, and gives Astra a 99% probability of being the more capable of the two.
The intervals of all six priced models overlap: Claude Sonnet 5.5’s upper bound of 169.23 sits above Astra’s lower bound of 166.16, by our arithmetic. Epoch’s point is that overlapping intervals no longer stop a direct comparison.
Which Astra
Epoch’s chart and data name GPT-6 Astra, with a date of 3 September 2026 in its scores file. On X, the account Haider quoted Epoch’s post as being about “Astra 6.2” and wrote: “i guess that it ends up at 173 ECI.” That is Haider’s own guess, 3.93 points above GPT-6 Astra’s score, by our arithmetic.
A model labelled gpt-6.2-astra appeared in Codex usage screenshots on X on 1 November, as we reported. Epoch’s scores file, as we downloaded it on 5 November, lists one Astra model: GPT-6 Astra.
Connecting the Coverage
Our related pieces on frontier model rankings and prices:
As we read Epoch’s numbers, the top three models sit within 0.71 ECI points of each other on point estimates, by our arithmetic, while their list prices differ by five times.
As we read the chart, the same-priced pair is the clearest gap: Claude Fable 5.1 lists at Astra’s $10 and $50, and Epoch gives Astra a 99% probability of being more capable.
As we read the new method, a probability says how likely one model is ahead, not by how much; a 72% figure leaves a 28% chance that Opus 5.5 is the more capable model, by our arithmetic.
Epoch AI, on X
“With better uncertainty estimates and a joint distribution over all models’ ECIs, we can now compare models directly.”
Three Implications
FOR BUYERS Epoch’s figures measure capability, not cost; the list prices above are the companies’ own, and contract prices can differ.
FOR MODEL COMPARISONS Epoch says the update lets it compare models “even when their individual ECI uncertainty intervals overlap.”
FOR GPT-6.2 ASTRA Epoch’s scores file lists one Astra model, GPT-6 Astra; a 173 ECI figure for a newer Astra is Haider’s guess.
The framework says: “The new job of finance is to manage the ratio between the two — work produced per token burned.”
As we read it, Epoch’s probabilities and the list prices are two halves of that ratio for model choice: how likely a model is ahead, and what each token costs. The framework does not discuss Epoch or these models.
What Is Not Established
The probabilities are from Epoch’s chart, built on data as of 26 October; the point scores are from its file labelled 5 November, so the two may differ slightly. ECI is one composite of many benchmarks, and Epoch’s figures measure capability, not cost or speed on a given task. List prices are not what large customers pay under contract.
This Week In The AI Supercycle
Anthropic’s Reality Check, The Capital Squeeze, & The New World Order
This week’s issue: for two years the AI build-out was priced as a growthstory. This week it started being priced as a credit story.
Epoch’s new method turns overlapping scores into odds. On its numbers, the top of the market is close: GPT-6 Astra, at 2.5 times Opus 5.5’s list price, has a 72% chance of being the more capable model, and a 54% chance against GPT-6.1 Sol at a fifth of the price.
A note on sourcing. Epoch AI’s figures are from its 4 November posts on X, the chart attached to them (data as of 26 October 2026) and its scores file labelled as updated on 5 November. List prices are from OpenAI’s and Anthropic’s pricing pages as recorded in our 4 November piece. The gaps and price ratios 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.
