September gave us the two models most engineering teams are now choosing between: Anthropic's Claude Opus 5.5 (September 22) and OpenAI's GPT-6 Astra (September 3). Both are excellent. They are excellent in different ways, and only one of them can be bought by absolutely everyone today.

Coding and context

Opus 5.5 leads the Opus tier and posts the strongest agentic coding results we have tracked this cycle, with a 1M-token context that swallows monorepos whole. For multi-file refactors and long-running agent loops, it is the model our testing kept coming back to.

Astra's ceiling on the hardest single-shot reasoning is the highest we have seen, consistent with Muse Wonder's preview scores being described as 'level with GPT-6 Astra' at ~52.4 on the Artificial Analysis index. But Astra began life limited to selected organizations, with 'GPT-6 Pro' surfacing inside ChatGPT Pro, and the staged October rollout is still working through Plus, Pro, Business and Enterprise plans.

Verdict: Opus 5.5 is the coding buy today. Astra is the ceiling you benchmark against while your access request is pending.

The money

Opus 5.5 lists at $4/$20 per million tokens. Astra lists at $10/$50, two and a half times as much, and the same bracket as Anthropic's own Mythos-class Fable 5.1. At those numbers, Astra needs to win your evals by a margin that justifies 2.5x, not by a nose.

Full scores for both models, alongside Gemini 4 Argon, Grok 4.7 and the open-weight field, are in our Reviews & Benchmarks comparison.