OpenAI Launches GPT-6 Astra Frontier AI Model
OpenAI has introduced GPT-6 Astra, rolling out across ChatGPT, the OpenAI API, and AWS with pricing set at $10 per million input tokens and $50 per million output tokens. The new frontier model achieves a landmark 99.9% score on the ARC-AGI 3 benchmark when utilizing a specialized provider adapter harness. GPT-6 Astra directly challenges Anthropic's Claude Fable 5 by matching its API pricing while delivering top-tier performance on cybersecurity tasks and near-perfect retrieval accuracy on long contexts up to 1 million tokens. Its launch marks substantial progress in long-context stability and cost efficiency for autonomous coding agents. Although Astra scored 99.9% on ARC-AGI 3 using a custom harness that preserves opaque reasoning states between requests, it scored 62.7% when tested with the default harness. Additionally, testing by Artificial Analysis shows Astra tying GPT-5.6 Sol with a score of 61 on their Intelligence Index, slightly behind Claude Fable 5.1.
## BACKGROUND
The ARC-AGI benchmark evaluates an AI's fluid intelligence and abstract reasoning capability through novel interactive environments and grid puzzles, serving as a key measure for progress toward artificial general intelligence. Meanwhile, Anthropic's Claude Fable 5 is a leading frontier AI model built for complex reasoning, long-running software engineering tasks, and autonomous project execution.