OpenAI launched GPT-6 Astra on Thursday, September 3, 2026, calling it the company's "most intelligent and aligned model" and positioning it as a computer-use agent rather than a conventional chatbot2,6,8. OpenAI President Greg Brockman ended a press briefing with the declaration: "Welcome to the AGI era".
The model's headline benchmark numbers are large but carry significant caveats. OpenAI reports Astra scored 98.6% on ARC-AGI-3, the interactive reasoning benchmark that stumped frontier models at sub-1% when it launched in March5,7. That score, however, was achieved through OpenAI's Provider Adapter harness, which preserves opaque reasoning state between requests and compresses longer conversations. ARC Prize ran the same model through its own standard harness and recorded 62.7%, a 36-point gap produced entirely by the software surrounding the model. The adapter-assisted run also cost less: $17,332 versus $26,098 for the standard harness.
On cybersecurity evaluations, Astra scored 100% on ExploitBench (up from 78.5% for GPT-5.6 Sol), 42.4% on ExploitGym (Sol: 30.3%), and 99.2% within four attempts on SRE-Bench binary reverse engineering (Sol: 68.7%)10. OpenAI designated Astra as the first model to reach its "Critical" cybersecurity threshold under its Preparedness Framework. The company said Astra's ability to identify and develop zero-day exploits can help defenders find and patch weaknesses, and that it found two previously unknown vulnerabilities during evaluation against recent V8 bugs.
Independent evaluations paint a more mixed picture on general intelligence. Artificial Analysis scored Astra at 61 on its Intelligence Index, tying GPT-5.6 Sol and trailing Anthropic's Claude Fable 5.1 at 6615. On the Artificial Analysis Coding Agent Index, Astra scored 67, matching Claude Opus 5, Fable 5, and Meta's Muse Spark 1.3, but falling short of Fable 5.1's leading score of 70. Analytics Insight characterized the model as showing "flat general intelligence" while leading "sharply in computer use and long coding sessions"14.
Other reported benchmarks include 97.6% on FrontierMath Tier 4, 72.6% on an offline subset of OSWorld 2.0 (Sol: 65.7%), 96% on GPQA Diamond, 74.1% on DeepSWE v1.1, and 95.9% on BenchCAD. Early community-run tests returned lower numbers on some tasks: 13% on Mazebench without tools12 and 65.6% on ClockBench13.
Astra's API is priced at $10 per million input tokens and $50 per million output tokens, matching Claude Fable 5 and 5.14. The model supports a 1.05-million-token context window and up to 128,000 output tokens. Astra costs roughly 2.5 times more per token than GPT-5.6 Sol, though Artificial Analysis found it runs at under half the cost per finished coding task compared to Fable 5 for the same score.
The rollout itself was rocky. Hours after the announcement, CEO Sam Altman apologized for what he called "a messy rollout," acknowledging that broad access had not yet begun for API customers or ChatGPT subscribers. Altman said OpenAI expected to begin broader access "in the near future," starting with Pro subscribers. OpenAI's Codex engineering lead Thibault Sottiaux said the rollout would take a few days to complete. Initial access went to organizations in OpenAI's Daybreak cybersecurity program, with Plus, Pro, Business, Enterprise, API, AWS Bedrock, and Microsoft Azure users expected to follow.
Brockman framed AGI as no longer tied to a contractual trigger from OpenAI's earlier agreement with Microsoft, describing it instead as a "mission concept" or "spiritual concept". OpenAI said Astra was trained using more than 100,000 GPUs at its Stargate site in Texas and is the first OpenAI model where earlier models played a significant role in supervising training.
ANALYSIS The ARC-AGI-3 gap between 62.7% and 98.6% depending on harness configuration underscores that benchmark scores increasingly measure the system stack, not the model alone. For practitioners evaluating Astra against competitors, the harness and serving configuration may matter as much as the underlying weights.