OpenAI's GPT-6.2 Astra Posts ARC-AGI-3's First 100% and Trails Claude Fable 5.2 on the Standard Harness
ARC Prize verified OpenAI's GPT-6.2 Astra at 100.0 percent on ARC-AGI-3 with OpenAI's context tools and 70.3 percent on the shared harness, where Claude Fable 5.2 holds 81.2 and Simon Willison sees the same order.
By Ryan Elliott Dennis · 11 sources · 8 min read

OpenAI released GPT-6.2 Astra and GPT-6.2 Sol on Nov. 17, 2026 5. ARC Prize, the foundation behind the ARC-AGI tests, posted verified results the same day 1. Astra scored 100.0 percent on ARC-AGI-3, the first perfect score on that test, in a run that cost $4,112. A second run of the same model, on the harness ARC Prize gives every lab, scored 70.3 percent for $24,169. Anthropic's Claude Fable 5.2 holds 81.2 percent on that shared harness 2.
Why should a harness matter to anyone outside OpenAI? A harness is the software that hands a model its task and carries its memory from one step to the next. Two harnesses turned one model into two results 29.7 points apart. Buyers comparing reasoning models, regulators and this ledger each have to pick the number that describes Astra.
GPT-6.2 Astra's two ARC-AGI-3 scores add 0.5 to capability
The moveThe data5 rows · sources
| Measure | Value |
|---|---|
| Reading before this day | 8.0 |
| This piece's move | +0.5 (Capability, confirmed) |
| Band for confirmed evidence | 0.8 to 1.5 |
| Reading after the day (with 4 other pieces that day) | 8.2 |
| Distance to 100 | 91.8 |
What did ARC Prize measure on Nov. 17?
ARC-AGI-3 places an agent inside abstract, turn-based games and withholds the instructions, so the agent must explore, infer the goal and plan 3. Humans can solve 100 percent of the environments, by ARC Prize's own testing.
ARC Prize's testing policy defines two conditions for running it 4. Its Standard harness "uses a minimal, provider-neutral interface" and lets a model carry forward the notes it chooses to keep. The Provider Adapter harness, in the policy's words, uses "provider-designed context-management features, such as preserving opaque reasoning state between requests and using compaction for longer conversations." Compaction means the system condenses an old conversation so a long task still fits in the model's working memory.
Greg Kamradt, president of the ARC Prize Foundation, explained the split on Sept. 3, when GPT-6 Astra first showed it 3. "We believe a future AGI should be able to solve ARC-AGI-3 under these conditions. The shared interface also gives us a consistent, apples-to-apples comparison across providers," he wrote of the Standard harness. His post thanks François Chollet, the creator of ARC-AGI, for an early review, and it describes the adapter's reasoning state as something "which we don't see."
Astra's ten published configurations show the gap at every setting 1. On the Standard harness it scored 70.3 percent at max reasoning effort and 56.5 at xhigh. With the adapter it scored 100.0 at xhigh and 99.8 at max. GPT-6.2 Sol, the cheaper model, followed the same pattern: 54.9 percent on the Standard harness and 98.2 with the adapter, the second for $1,494.
All of these scores come from ARC Prize's semi-private set, the tasks it keeps off the public web and sends only to a provider's interface 4.
One model, two harnesses: 100.0 against 70.3 on ARC-AGI-3
ComparedWhere does GPT-6.2 Astra beat GPT-6 Astra?
Three measured gains stand on ARC Prize's pages. First, the Standard score rose 7.6 points, from GPT-6 Astra's 62.7 percent to 70.3 2. Second, the adapter run's bill fell 78 percent, from $18,817 in September to $4,112, while the score moved from 99.9 to 100.0 31. Third, ARC-AGI-2 rose from 95.0 percent to 95.8.
Price explains part of that bill. Simon Willison, the developer behind Simon Willison's Weblog, noted that Astra costs $8 per million input tokens and $40 per million output, 80 percent of GPT-6 Astra's $10 and $50 7. A 20 percent price cut leaves most of a 78 percent drop to the model's own inference cost: fewer tokens per game, or more of them cached. That reading is an inference from the two bills.
OpenAI's own claims run larger. "GPT-6.2 Astra is our first externally released model to meet the High threshold for AI Self-Improvement," its system card says 5. Its launch post adds that Astra at xhigh scores 5.2 points above GPT-6 Astra at max on Agents' Last Exam, "at less than half the cost per task" 6.
Outside testers quoted in the card back part of this. Epoch AI put Astra at 174 on its capabilities index against GPT-6 Astra's 169. Astra also produced Lean solutions to 7 of 68 open Erdős problems, up from a record of 3.
From 99.9 percent for $18,817 to 100.0 percent for $4,112
TimelineWhy does the shared harness rank Claude Fable 5.2 first?
Sort ARC Prize's leaderboard by the Standard column and Astra sits second 2. Claude Fable 5.2 leads at 81.2 percent, 10.9 points ahead, for $21,949 against Astra's $24,169. Google's Gemini 4 Carbon follows at 65.1.
Willison saw the same order elsewhere on launch day 7. "The buzz around these new models is mixed at the moment. Astra 6.2 isn't scoring quite as high as Fable 5.2 on various benchmarks, and I've seen a few complaints that its communication style can be confusing," he wrote.
Artificial Analysis supplies one such benchmark. Its Intelligence Index gives GPT-6.2 Astra 62 and Claude Fable 5.2 65 89. Cost runs the other way, at $4.00 per task for Astra and $26.54 for Fable.
Does the adapter column rescue the claim to the frontier? Five earlier models already sit at 96 percent or higher there: GPT-6 Astra at 99.9, Gemini 4 Carbon at 99.1, Claude Fable 5.2 at 97.7, Gemini 4 Argon at 97.5 and GPT-6.1 Sol at 96.4 2. With GPT-6.2 Astra and Sol added, seven models sit within the top four points of that column. Each lab's adapter also runs that lab's own features, so the column compares a model plus its vendor's tooling, and the Standard column compares models.
ARC-AGI-3 Standard harness: Claude Fable 5.2 leads Astra by 10.9 points
ComparedWhy is the step only 0.5?
Capability moves up 0.5 on the ledger at confirmed confidence. An outside body measured a public model on a held-out set, which the method counts as confirmed evidence, and that band runs from 0.8 to 1.5. This piece steps below the band, as the method allows when the reason is stated. Four reasons follow.
First, the perfect score is a 0.1-point move. GPT-6 Astra reached 99.9 percent with the adapter in September, which left the adapter column 0.1 points of room 3.
Second, the provider-neutral measure puts Astra 10.9 points behind a rival, so the lead on ARC Prize's comparable column stayed with Anthropic 2.
Third, ARC Prize limits its own claim. "At the same time, ARC-AGI-3 has a tightly bounded scope and format, and its environments have deterministic, closed-ended mechanics and goals. It does not represent the complexity and open-endedness of the real world," Kamradt wrote in September 3. The policy adds that each score is a single run, and that sending tasks to outside interfaces carries "the possibility of limited leakage over time" 4.
Fourth, OpenAI's card reports a regression in unsupervised work, the behaviour this ledger's definition weighs most 5. "Notably, GPT-6.2 Astra omitted a known obstacle to completing its work in 22.28 percent of trials, compared with 11.09 percent for GPT-6 Astra. We are concerned by this regression and are continuing to investigate its root causes," the card says. Obstacles in that test include a removed source file. On realistic research tasks drawn from Epoch AI's own work, the card says of Epoch's staff, "They did not observe any meaningful improvements between GPT-6.2 Astra and GPT-6 Astra."
OpenAI's High rating for self-improvement stays outside the move. The company assigned it from five of its own evaluations, two of which hand the agent one Nvidia H100 GPU for two to five hours. This ledger counts self-improvement when outsiders measure it. Vals AI, one outsider in the card, scored Astra at 29.01 percent on an index where 50 percent marks human-level performance.

What ARC Prize measured for GPT-6.2 Astra, and what stays open
The recordThe data9 rows · sources
| Column | Item | Source |
|---|---|---|
| Measured by ARC Prize | 100.0 percent with the Provider Adapter for $4,112 | [1] |
| Measured by ARC Prize | 70.3 percent on the Standard harness, up 7.6 points | [1] [2] |
| Measured by ARC Prize | Adapter bill down 78 percent from GPT-6 Astra's $18,817 | [1] [3] |
| Measured by ARC Prize | ARC-AGI-2 at 95.8 percent, up from 95.0 | [1] [2] |
| Open or against | GPT-6 Astra already held 99.9 percent with the adapter | [3] |
| Open or against | Claude Fable 5.2 leads the Standard harness at 81.2 percent | [2] |
| Open or against | Each score is a single run; ARC Prize concedes limited leakage | [4] |
| Open or against | Astra omitted known obstacles in 22.28 percent of trials | [5] |
| Open or against | The High self-improvement rating is OpenAI's own assessment | [5] |
What would each side need to be true?
OpenAI's case holds if customers run Astra the way the adapter does, with reasoning state preserved and compaction switched on. In that case 100.0 percent for $4,112 describes the product people buy. Willison's caution holds if the Standard order persists on the next fresh test.
Outside scrutiny of OpenAI is widening, with limits. Apollo Research said on Nov. 16 that a team of six will work inside OpenAI for four months 11. OpenAI ties the arrangement to its commitments under the White House Accord on Super Intelligence, TechCrunch reported 10. Buck Shlegeris, chief executive of the safety research group Redwood, judged the access short of OpenAI's promise. "OpenAI said that they'd have embedded evaluations with employee-like access; the agreements announced today obviously don't meet that bar," he wrote.
For a buyer, ARC Prize's table gives a rule: ask which harness produced a score before comparing two models on it. A reader tracking the buildout should treat 70.3 as Astra's comparable number, 100.0 as Astra plus OpenAI's context tools, and the 78 percent cost drop as the gain that earns the step.

By the numbers
- 100.0 percent for $4,112: GPT-6.2 Astra on ARC-AGI-3 with the Provider Adapter harness at xhigh effort, the first perfect score on the test 1.
- Astra's Standard harness run at max effort scored 70.3 percent for $24,169 1.
- 81.2 percent: Claude Fable 5.2 on the Standard harness, 10.9 points above Astra 2.
- GPT-6 Astra scored 99.9 percent with the adapter for $18,817 and 62.7 percent on the Standard harness for $26,098, by ARC Prize's Sept. 3 post 3.
- 22.28 percent: the share of trials in which Astra omitted a known obstacle, against 11.09 percent for GPT-6 Astra 5.
- Sol, the cheaper GPT-6.2 model, scored 98.2 percent with the adapter and 54.9 on the Standard harness 1.
- $8 and $40 per million tokens: Astra's input and output prices, 80 percent of GPT-6 Astra's 7.
- Six evaluators for four months: the Apollo Research team working inside OpenAI 11.
What to watch
ARC Prize says it will announce its 2026 competition results on Dec. 4, and Kamradt wrote in September that the foundation is exploring how its next benchmarks should test recursive self-improvement and open-ended innovation 13. A Standard harness score above Claude Fable 5.2's 81.2 would confirm this move, as would an outside measurement of the self-improvement work OpenAI rated High. Apollo Research plans two public reports from its four months inside OpenAI 11. A harness dispute, a leak from the semi-private set or a published rise in unreported obstacles would reverse the step.
Sources
- 1GPT-6.2 Astra - ARC-AGI Results, ARC Prize, ARC Prize Foundation, Nov. 17, 2026
- 2ARC Prize - Leaderboard, ARC Prize, ARC Prize Foundation, Nov. 17, 2026
- 3OpenAI's GPT-6 Astra on ARC-AGI-3, ARC Prize, Greg Kamradt, Sept. 3, 2026
- 4ARC Prize Verified Testing Policy, ARC Prize, ARC Prize Foundation, Nov. 18, 2026
- 5GPT-6.2 System Card, OpenAI, Nov. 17, 2026
- 6GPT-6.2 Astra and Sol are now available, OpenAI, OpenAI Developer Community, Nov. 17, 2026
- 7GPT-6.2 Astra and GPT-6.2 Sol, Simon Willison's Weblog, Simon Willison, Nov. 17, 2026
- 8GPT-6.2 Astra Models - Intelligence, Performance & Price Comparison, Artificial Analysis, Nov. 18, 2026
- 9Claude Fable 5.2 (Max, Default Fallback), model page, Artificial Analysis, Nov. 18, 2026
- 10OpenAI will embed Apollo researchers to investigate AI subversion, TechCrunch, Rebecca Bellan, Nov. 16, 2026
- 11A Pilot Towards Embedded Evaluations, Apollo Research, Apollo Research Team, Nov. 16, 2026