OpenAI Starts Rolling Out GPT-6 Astra With 99.9% on ARC-AGI 3

OpenAI has begun rolling out GPT-6 Astra. The model went live today for a limited set of organizations and will reach all ChatGPT Plus, Pro, Business and Enterprise users in the coming days, along with the OpenAI API and AWS. The API label will be gpt-6-astra.

Pricing is $10 per million input tokens and $50 per million output tokens, the same rate as Claude Fable 5 and 5.1. Simon Willison describes Astra as OpenAI's answer to Fable, noting it scores higher than Fable on most of OpenAI's self-reported benchmarks.

The ARC-AGI 3 number depends on the harness

Astra scored 99.9% on ARC-AGI 3, released in March. The ARC-AGI blog notes that result came from OpenAI's custom "Provider Adapter" harness at a cost of $19K, while the default ARC-AGI harness produced 62.7% for $26K. The Provider Adapter preserves opaque reasoning state between requests and applies compaction on longer conversations so the model can reuse earlier work. Fable 5 has no published result on this benchmark yet.

On security tasks Astra hit 100% on ExploitBench, against 78.5% for GPT-5.6 Sol, 42.4% on ExploitGym versus Sol's 30.3%, and 99.2% within four attempts on SRE-Bench binary reverse engineering versus Sol's 68.7%. On OpenAI's eight-needle long context benchmark it recorded 100% between 256K and 512K tokens and 96.3% between 512K and 1M.

Level with Sol on intelligence

Artificial Analysis places Astra at 61 on its Intelligence Index, equal to GPT-5.6 Sol and 5 points below Claude Fable 5.1 (max with fallback). It also trails Meta's newly released Muse Spark 1.3 (max). On the Coding Agent Index, the same analysis says Astra leads the cost efficiency frontier: at max effort it costs roughly the same as Sol while scoring 2 points higher, and per task it runs at less than half the cost of Claude Fable 5 for the same score.

Willison says he has not tested the model yet and will write more once he has access. Full post at Simon Willison's Weblog.