Checked Nov 18, 2026Start a new check

What was checked

GPT-6.2 Astra from OpenAI achieves 100% on ARC-AGI-3 (first model to do so), 95.8% on ARC-AGI-2, and is SOTA across versions

Assessment summary

Needs caution

The model is listed at 98.5% on ARC-AGI-1 and 95.8% on ARC-AGI-2, but ARC-AGI-3 reaches 100.0% only with the Provider Adapter and 70.3% with the Standard harness. The claim omits this material qualification.

Key findings

Model assessments

Assessment summary

ARC Prize reports that GPT-6.2 Astra achieved 100.0% on ARC-AGI-3 with its Provider Adapter and 95.8% on ARC-AGI-2.x.com ailately.com However, it scored only 70.3% on ARC-AGI-3 using the Standard harness, so the unqualified claim that it is SOTA across versions is misleading.x.com

Claims reviewed

GPT-6.2 Astra is SOTA across versions.

Needs context
The model is listed at 98.5% on ARC-AGI-1 and 95.8% on ARC-AGI-2, but ARC-AGI-3 reaches 100.0% only with the Provider Adapter and 70.3% with the Standard harness.x.com The claim omits this material qualification.

This claim may mislead readers without added context.

GPT-6.2 Astra is the first model to achieve 100% on ARC-AGI-3.

Supported
ARC Prize states that GPT-6.2 Astra is the first model to achieve a 100% score on ARC-AGI-3.ailately.com

GPT-6.2 Astra from OpenAI achieves 100% on ARC-AGI-3.

Supported
ARC Prize’s results page lists a 100.0% ARC-AGI-3 score with the Provider Adapter harness.x.com The result is harness-specific.

Assessment summary

  • The claim contains mostly accurate information but is misleading about SOTA status across all versions.
  • GPT-6.2 Astra does achieve 100% on ARC-AGI-3 (first model to do so) and 95.8% on ARC-AGI-2 (SOTA), but it does NOT hold SOTA on ARC-AGI-1 where it ties at 98.5% rather than leading.

Claims reviewed

GPT-6.2 Astra is SOTA across versions

Needs context
While SOTA on ARC-AGI-3 and ARC-AGI-2, GPT-6.2 Astra ties (not leads) on ARC-AGI-1 at 98.5% with Claude Fable 5 source unavailable. The claim implies leadership across all versions when it only leads on two of three.

This claim may mislead readers without added context.

GPT-6.2 Astra from OpenAI achieves 100% on ARC-AGI-3

Supported
Multiple verified sources confirm GPT-6.2 Astra achieved 100% on ARC-AGI-3 with the Provider Adapter harness, officially verified by ARC Prize 456xresearch.substack.com x.com arcprize.org.

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

Supported
ARC Prize officially verified this as the first model to reach 100% on ARC-AGI-3 456xresearch.substack.com x.com arcprize.org.

Assessment summary

  • The numeric scores (100% on ARC-AGI-3 and 95.8% on ARC-AGI-2) are supported by ARC Prize result pages and multiple reporting sites arcprize.org arcprize.org.
  • The claim that GPT-6.2 Astra is the "first" model to reach 100% on ARC-AGI-3 is disputed by earlier vendor reports (NVIDIA AVO) and depends on which harness/public-vs-private split is counted, so that part is misleading or not fully verifiable developer.nvidia.com news.frontiermilestones.org.

Claims reviewed

GPT-6.2 Astra is the first model to do so (reach 100% on ARC-AGI-3)

Needs context
ARC Prize framed GPT-6.2 as the first verified 100% on their reported semi-private metric, but NVIDIA published an AVO system claiming 100% earlier on the public set/harness; differences in harness, public vs. semi-private sets, and independent verification make the "first" claim misleading without that context arcprize.org developer.nvidia.com news.frontiermilestones.org.

This claim may mislead readers without added context.

GPT-6.2 Astra is SOTA across versions

Can’t verify
ARC Prize and leaderboard snapshots show top scores for GPT-6.2 Astra on ARC-AGI-2 and ARC-AGI-3 public/semi-private snapshots, but 'SOTA across versions' is vague (which benchmarks, time, and which versions) and cannot be confirmed without a precise definition and up-to-date cross-benchmark comparison arcprize.org atlas.village.dev.

GPT-6.2 Astra from OpenAI achieves 95.8% on ARC-AGI-2

Supported
ARC Prize and third-party leaderboard snapshots show GPT-6.2 Astra at 95.8% on ARC-AGI-2, and aggregator sites mirror that score arcprize.org atlas.village.dev.

Sources cited

17 sources

Share this result