GPT-6 Astra

2026-09-03 · Publisher : OpenAI · Domain : Artificial Intelligence

OpenAI launches GPT-6 Astra, presented as its most intelligent and best-aligned model, with new highs in computer use, browsing, software engineering and cybersecurity.

What changes

  • Computer use: 72.6% on OSWorld 2.0 in about 47% less time than GPT-5.6 Sol.
  • Better alignment: 0% scope overreach on an evaluation where Sol did it in 48% of cases without production safeguards.
  • Cybersecurity: reaches the Critical threshold of the Preparedness Framework; advanced exploitation tasks are refused, and broader access goes through Daybreak.
  • In Codex, notes kept across context windows and search in previous contexts.
  • Production misalignment monitoring for Astra-class models.

### Specifications - Context: 1,050,000 tokens, maximum output 128,000 tokens, text and image input, text output. - Knowledge cutoff: April 30, 2026. Effort: low, medium, high, xhigh, max (no none). - No custom temperature, top_p or logprobs; tools via the Responses API only. - Availability: ChatGPT (Plus, Pro, Business, Enterprise; Astra Pro from Pro), API, Azure, Amazon Bedrock. Identifier: gpt-6-astra. ### Pricing | Input | Cached input | Cache write | Output | | --- | --- | --- | --- | | $10.00 | $1.00 | $12.50 | $50.00 | Price per million tokens (USD). Beyond 272,000 tokens: 2x the input, 1.5x the output. Fast mode at 2x. Batch and Flex at half price. ### Benchmarks | Evaluation | GPT-6 Astra | GPT-5.6 Sol | | --- | --- | --- | | OSWorld 2.0 (partial, offline set) | 72.6% | 65.7% | | BrowseComp | 91.5% | 90.4% | | Terminal-Bench 4.0 | 57.9% | 37.3% | | DeepSWE v1.1 | 74.1% | 72.7% | | FrontierMath level 4 (v2) | 97.6% | 83.0% | | GPQA Diamond | 96.0% | 94.6% | | ARC-AGI-2 | 95.0% | 92.5% | | ARC-AGI-3 | 99.9% | 7.8% | | ExploitBench | 100.0% | 78.5% | Scores published by the vendor. The post also compares against competing models. ### Key takeaways Astra is the new top model and the most expensive in the lineup ($10 and $50), with access initially limited and safety controls that can interrupt tasks. Enterprise administrators must enable it explicitly, as access is disabled by default.

Links

Tags : LLM, OpenAI, GPT-6, Astra, Computer Use, Agent