OpenAI’s GPT-6.2 Astra scores 174 on the Epoch Capabilities Index, “on par with Fable 5.2”, Epoch AI posted on X at 18:42 UTC on Tuesday, 17 November, the day OpenAI released the model. Epoch says the result is “based on testing we did prior to launch using only our own benchmarks” and that the values “may change in the future”.
Our number: in Epoch’s data file, GPT-6.2 Astra scores 174.17 and Anthropic’s Claude Fable 5.2 scores 173.5, a gap of 0.67 points, and each score sits inside the other model’s confidence interval, by our arithmetic. The two models also carry the same list price, $8 per million input tokens and $40 per million output tokens, as we reported earlier on Tuesday.
Business Pill · Usage-Based Pricing
An evergreen Business Engineer explainer on usage-based pricing, paying for work rather than seats. It teaches the general idea only and does not discuss OpenAI, Anthropic or Epoch AI.
The key insight: On Epoch AI’s capability index, OpenAI’s and Anthropic’s newest flagships sit 0.67 points apart, inside each other’s confidence intervals, at the same $8 input and $40 output list price, by our arithmetic.
What Epoch Reported
Epoch says GPT-6.2 Astra joins Fable 5.2 “above what the pre-Astra trend for reasoning models would have predicted”. On 9 November, when Anthropic released Fable 5.2, Epoch said it “sets a new record on the Epoch Capabilities Index (ECI) at 174”, as we reported then.
Epoch’s data file lists GPT-6.2 Astra’s interval as 170.29 to 178.58 and Fable 5.2’s as 170.09 to 177.36. Google DeepMind’s Gemini 4 Carbon, released 10 November, is third at 169.78, 4.39 points below GPT-6.2 Astra, by our arithmetic.

How Far Each Lab Moved
GPT-6.2 Astra is 4.97 points above GPT-6 Astra, which Epoch’s current file scores at 169.2 and dates to 3 September, 75 days earlier, by our arithmetic. In the same file, Claude Fable 5.2 is 6.41 points above Claude Fable 5.1 (167.09, dated 1 September), 69 days apart.
Gemini 4 Carbon is 3.21 points above Gemini 4 Argon (166.57, dated 13 October), 28 days apart. A year earlier, the highest score in Epoch’s file for a model released before 17 November 2025 was GPT-5 Pro’s 151.27; GPT-6.2 Astra is 22.9 points above it, by our arithmetic.
What a Point on the Index Means
Epoch says “Absolute ECI values are meaningless by themselves, but meaningful comparisons can be made between models”. The scale is calibrated so that Claude 3.5 Sonnet scores 130 and GPT-5 scores 150, and Epoch says these calibration points “may be revisited”.
Epoch also says that “at the time of ECI’s launch, a 5 point gain in ECI appeared to roughly correspond to a doubling of the METR Time Horizon”. GPT-6.2 Astra’s 4.97-point step over GPT-6 Astra is close to that size; Epoch did not publish a METR estimate for the new model.
The Business Lever
The lever is API pricing. On Epoch’s index the two highest-scoring models now sit 0.67 points apart at the same $8 and $40 list price, so for developers choosing between OpenAI and Anthropic at the top of the range, list price does not separate them.
OpenAI also says GPT-6.2 Astra is its first publicly released model rated High for AI self-improvement, and it released Astra for Law, a legal configuration of the same model, on the same day.
The Structural Read
As we read Epoch’s data, the 0.67-point gap is smaller than the uncertainty Epoch attaches to either score, so the index does not rank one model above the other.
As we read the step sizes, OpenAI’s 4.97-point gain over GPT-6 Astra took 75 days and Anthropic’s 6.41-point gain over Fable 5.1 took 69 days, by our arithmetic from Epoch’s file.
Epoch AI, 17 November 2026
“This is based on testing we did prior to launch using only our own benchmarks”
Three Implications
FOR API BUYERS GPT-6.2 Astra and Claude Fable 5.2 list at $8 per million input tokens and $40 per million output tokens, and Epoch scores them 174.17 and 173.5.
FOR OPENAI Epoch says GPT-6.2 Astra joins Fable 5.2 above what the pre-Astra trend for reasoning models would have predicted; its score is from pre-launch testing and may change.
FOR GOOGLE Gemini 4 Carbon, released 10 November, scores 169.78 in Epoch’s file, 4.39 points below GPT-6.2 Astra, by our arithmetic.
The Business Engineer Lens
This story maps onto the Business Engineer framework The Agentic Harness War.
The framework says: “Foundation models become relatively commoditized as orchestration layers capture differentiation.”
As we read it, two flagships at the same list price and within each other’s error range on Epoch’s index fit that description at the model layer. The framework predates this release and does not discuss GPT-6.2 Astra or Epoch’s index.
What Is Not Established
The 174 score comes from Epoch’s pre-launch testing on its own benchmarks only, and Epoch says the index is continually updated as more results arrive. Fable 5.2’s score has already moved from the 174 Epoch posted on 9 November to 173.5 in the current file.
The confidence intervals are Epoch’s; the gaps, day counts and year-on-year comparison are our arithmetic. We did not contact Epoch AI, OpenAI or Anthropic.
The Bottom Line
Epoch AI scores GPT-6.2 Astra at 174.17 and Claude Fable 5.2 at 173.5, a 0.67-point gap inside both models’ intervals, at the same $8 and $40 list price, by our arithmetic. GPT-6.2 Astra is 4.97 points above GPT-6 Astra, 75 days on.
95,000+ executives read Business Engineer for the AI strategy frameworks cited by ChatGPT, Claude, and Perplexity.
A note on sourcing. The scores, confidence intervals and release dates are from Epoch AI’s Epoch Capabilities Index data file, downloaded on 17 November 2026; the GPT-6.2 Astra result is from Epoch’s pre-launch testing on its own benchmarks, as Epoch said on X. Calibration and interpretation notes are from Epoch’s ECI page. List prices are from our earlier report on the GPT-6.2 release. The gaps, day counts and year-on-year comparison are our arithmetic. We did not contact Epoch AI, OpenAI or Anthropic.
Sources: Epoch AI on X, 17 Nov 2026 · Epoch AI, Epoch Capabilities Index data (eci_scores.csv), downloaded 17 Nov 2026 · Epoch AI, Epoch Capabilities Index page









