“OpenAI launched GPT-6 Astra, a new AI model achieving 98.6% on ARC-AGI-3, 100% on ExploitBench, 97.6% on FrontierMath Tier 4 (v2), and 74.1% on DeepSWE v1.1, outperforming GPT-5.6 Sol and Claude Fable 5.1”

The release is real and fully documented on official channels. OpenAI published a launch page, a system card, a preparedness write-up and an API model page for GPT-6 Astra on September 3, 2026, and Microsoft published a Foundry availability post. This part of the claim needed no adjudication. All four numbers are real and traceable. None is fabricated. What the sources show is that each one carries a condition that the post removed. ARC-AGI-3. ARC Prize, the benchmark's operator, ran the model itself and published verified results. With ARC Prize's own provider-neutral Standard harness, Astra scored 62.7% at max reasoning, at a run cost near $26,000. With OpenAI's Provider Adapter harness, which ARC Prize describes as preserving opaque reasoning state between requests and using compaction, the model reached 99.9% at high reasoning. The 98.6% figure in the post is the same Provider Adapter condition at max reasoning. So 98.6% and 62.7% are the same model weights, same benchmark, two interfaces. ARC Prize said it will label the two conditions separately going forward. The New Stack reported that the harness producing the headline number is not itself part of the shipped product. Note that OpenAI's own launch page headlines 99.9%, not 98.6%, because its tables report the maximum score at any reasoning effort. ExploitBench. The 100% is confirmed in OpenAI's own preparedness document. That same document states OpenAI had contamination concerns about the benchmark, and therefore built a contamination-controlled internal port using high-severity V8 vulnerabilities disclosed between June and August 2026. On that controlled version, reporting of OpenAI's figures puts Astra at roughly 39.0% arbitrary code execution. OpenAI's own footnote further warns that some included vulnerabilities may not permit arbitrary code execution under the evaluation's constraints, so a 100% success rate may not be achievable in principle. FrontierMath Tier 4 (v2). The 97.6% is OpenAI's reported figure. Epoch AI's own benchmark page carries a standing disclosure that FrontierMath was developed with OpenAI funding and that OpenAI has exclusive access to a subset of the benchmark. Epoch also notes a June 2026 update that corrected errors in 42% of problems. No independent re-run of Astra on this benchmark was found. DeepSWE v1.1. This is the best-corroborated number in the claim. The independent Datacurve leaderboard lists Astra at roughly 74% with a stated uncertainty of ±3 in the shared mini-swe-agent setup, consistent with OpenAI's 74.1%. But the benchmark is 113 tasks, so one task is about 0.9 points, and the surrounding field is bunched: Claude Opus 5 at 73.7% and Gemini 3.8 Flash at 73.8%. Astra's margin over Opus 5 is 0.4 points, less than half a single task and far inside the published uncertainty band. Meta's Muse Spark 1.3 sits at 75.4% and was the leaderboard's number one on the claim's own publication date. On the comparative claim. OpenAI's own comparison table contains rows Astra loses. On Humanity's Last Exam with tools, Astra scores 57.2% against Claude Fable 5.1's 65.0%. On FrontierCode 1.1 splits, Fable 5 and Opus 5 edge Astra. Independently, Artificial Analysis measured Astra at about 61 on its Intelligence Index v4.1.1, level with GPT-5.6 Sol at 61 and behind Claude Fable 5.1 at roughly 66 and Opus 5 at roughly 63, while priced at 2.5 times Sol's rates. Artificial Analysis did find genuine gains for Astra in coding-agent performance and token efficiency. One contrary source deserves explicit handling. CryptoBriefing published a piece arguing the 98.6% claim was unverified and disconnected from a leaderboard topping out near 30%. The benchmark operator subsequently ran and published verified results confirming the figure under the stated harness. On this point the operator's own artifact governs, and that commentary is superseded.