OpenAI has released GPT-6 Astra, its new flagship vision-language model that achieves top or near-top rankings on major AI leaderboards while using significantly fewer tokens and costing less than competing models. The model is designed to meet OpenAI's critical cybersecurity level, with advanced cyber capabilities restricted to selected organizations.

  • Input/output supports up to 1,050,000 input tokens and 128,000 output tokens, with a knowledge cutoff of April 30, 2026.
  • Features include five reasoning levels, tool use (computer, shell, code interpreter), asynchronous tool calls, mid-turn steering, and context compaction.
  • GPT-6 Astra leads the ARC-AGI-3 leaderboard (62.7% with max reasoning) and Arena AI’s WebDev leaderboard, while ranking second on Artificial Analysis’ Intelligence Index v4.2.
  • Training utilized over 100,000 GPUs, incorporating alignment preferences and resistance to jailbreaks generated by the GPT-Red agent.
  • Pricing for API access is $10/$1/$12.50/$50 per million input/cached input/cache write/output tokens, with batch and flex options costing half the standard price.

The release demonstrates OpenAI's ability to produce high-performance models that are more cost-effective than previous state-of-the-art systems.