OpenAI’s GPT-6.2 Astra Is First to 100% on ARC-AGI-3: $4.1K

OpenAI’s GPT-6.2 Astra scored 100.0% on ARC-AGI-3, the ARC Prize Foundation said on X at 17:58 UTC on Tuesday, 17 November, in results it labels “Verified”. “GPT-6.2 Astra is the first model to achieve a 100% score on ARC-AGI-3”, ARC Prize wrote. The run cost $4.1K with ARC’s Provider Adapter harness; with ARC’s Standard harness the model scored 70.3% for $24.2K.

Our number: on ARC Prize’s leaderboard, the 100% run cost $4,112, against $18,817 for GPT-6 Astra’s 99.95% in September, about 4.6 times less, by our arithmetic. On the Standard harness, Anthropic’s Claude Fable 5.2 scores 81.17%, 10.86 points above GPT-6.2 Astra’s 70.31%, on the same leaderboard.

Business Pill · The Model and the Harness

An evergreen Business Engineer explainer on why products built on the same model differ. It teaches the general idea only and does not discuss OpenAI, ARC Prize or this result.

The key insight: ARC Prize verified GPT-6.2 Astra at 100% on ARC-AGI-3 with its Provider Adapter harness for a $4,112 run, about 4.6 times less than GPT-6 Astra’s 99.95% run, while on ARC’s Standard harness Claude Fable 5.2 scores 81.17% to Astra’s 70.31%, by our arithmetic.

What ARC Prize Reported

ARC Prize’s post lists three results for GPT-6.2 Astra: “ARC-AGI-3: 100.0% (SOTA), $4.1K (provider adapter), 70.3%, $24.2K (standard)”, 95.8% on ARC-AGI-2 at $0.50 per task and 98.5% on ARC-AGI-1 at $0.17 per task. The leaderboard rows are on ARC-AGI-3’s Semi-Private set.

The same leaderboard lists earlier models far lower on the Semi-Private set: Claude Opus 4.6 (Max) at 0.51% and GPT-5.4 (High) at 0.21%.

Best ARC-AGI-3 Semi-Private score per model on ARC Prize’s Standard harness vs its Provider Adapter harn
Best ARC-AGI-3 Semi-Private score per model on ARC Prize’s Standard harness vs its Provider Adapter harness: GPT-6.2 Astra 70.31% to 100%, Claude Fable 5.2 81.17% to 97.72%, Gemini 4 Carbon 65.12% to 99.09%, GPT-6 Astra 62.71% to 99.95%. Source: ARC Prize leaderboard, read 17 November 2026; costs are ARC’s evaluation-run costs. The point gains and the 4.6x cost ratio are our arithmetic.

Two Harnesses, Two Scores

ARC Prize says its Standard harness “enables a model to carry forward notes it chooses to keep with it throughout the environment”. It says the Provider Adapter harness “preserves opaque reasoning state between requests and uses compaction for longer conversations, allowing the model to reuse prior work”.

ARC Prize introduced the second harness with GPT-6 Astra in September and said then it would test all new models both ways. In its blog it says the Standard harness “gives us a consistent, apples-to-apples comparison across providers”, and that “We believe a future AGI should be able to solve ARC-AGI-3 under these conditions”.

Where the Other Labs Stand

On the Provider Adapter harness, the leaderboard lists Google’s Gemini 4 Carbon (Medium) at 99.09% for $4,560 and Claude Fable 5.2 (High) at 97.72% for $18,625. GPT-6.2 Astra’s High setting scored 99.80% for $3,878.

On the Standard harness, Claude Fable 5.2 leads at 81.17% (XHigh) for $21,949, and GPT-6.2 Astra (Max) is at 70.31% for $24,169, so Fable 5.2 scores higher for about 9% less run cost, by our arithmetic. GPT-6 Astra (Max) scored 62.71% on the same harness.

The Business Lever

The lever is the harness. On ARC-AGI-3 the same OpenAI model moves from 70.31% to 100% when it keeps its own hidden reasoning state between requests, a gap of 29.69 points, by our arithmetic; on the Standard harness, Anthropic’s model is ahead. We covered the same pattern in August, when Nvidia’s AVO harness took Claude Opus 5 to 100% on ARC-AGI-3.

The results land on the same day that an early tester published a letter criticising GPT-6.2 Astra, that Epoch AI scored it level with Claude Fable 5.2, and that OpenAI cut its price to $8 per million input tokens.

The Structural Read

As we read ARC Prize’s leaderboard, the harness moves the score more than the model does: the same GPT-6.2 Astra gains 29.69 points when it keeps its own hidden reasoning between requests, by our arithmetic.

As we read the costs, the September jump to 99.95% came at $18,817; the new 100% came at $4,112, so the change on this benchmark is mostly cost, not score.

ARC Prize, 17 November 2026

“GPT-6.2 Astra is the first model to achieve a 100% score on ARC-AGI-3”

Three Implications

FOR OPENAI ARC Prize lists GPT-6.2 Astra at 100% on the Provider Adapter harness and 70.31% on the Standard harness, both on the Semi-Private set.

FOR ANTHROPIC Claude Fable 5.2 leads the Standard harness at 81.17% and scores 97.72% on the Provider Adapter harness, according to the leaderboard.

FOR GOOGLE Gemini 4 Carbon scores 99.09% on the Provider Adapter harness for $4,560, according to the leaderboard.

The Business Engineer Lens

This story maps onto the Business Engineer framework The Agentic Harness War.

The framework says: “The surface, not the model, governs autonomy.”

As we read it, ARC Prize’s two harnesses show the same idea on a benchmark: one model, two scores, depending on how its working state is kept between steps. The framework predates this result and does not discuss ARC-AGI-3 or GPT-6.2 Astra.

What Is Not Established

ARC Prize’s cost figures are for its own evaluation runs; the leaderboard does not say how they map to a customer’s bill. The 4.6x comparison sets GPT-6.2 Astra’s XHigh run against GPT-6 Astra’s High run, the best score each posted on the Provider Adapter harness.

The leaderboard figures were read from arcprize.org at about 20:50 UTC and may be revised. We did not contact ARC Prize, OpenAI or Anthropic.

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 Agentic Harness War

A Business Engineer analysis of how value shifts from foundation models to the orchestration layers built on top of them.

Read the Map of AI →

The Bottom Line

ARC Prize verifies GPT-6.2 Astra as the first model to score 100% on ARC-AGI-3, with its Provider Adapter harness, for $4,112, about 4.6 times less than GPT-6 Astra’s 99.95% run, by our arithmetic. On ARC’s Standard harness the model scores 70.31%, behind Claude Fable 5.2’s 81.17%.

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

A note on sourcing. The GPT-6.2 Astra results and the Verified label are from the ARC Prize Foundation’s post on X of 17 November 2026; harness definitions are from ARC Prize’s blog; every other score and cost is from ARC Prize’s ARC-AGI-3 Semi-Private leaderboard, read at about 20:50 UTC. The point gaps and cost ratios are our arithmetic. We did not contact ARC Prize, OpenAI or Anthropic.

Sources: ARC Prize on X, 17 Nov 2026 · ARC Prize blog, OpenAI’s GPT-6 Astra on ARC-AGI-3 · ARC Prize, ARC-AGI-3 leaderboard, read 17 Nov 2026 · ARC Prize on X, 3 Sep 2026

Scroll to Top
FourWeekMBA