Claude Opus 5.5

Claude Opus 5.5 is Anthropic’s recommended model for most work from 22 September 2026, served as claude-opus-5-5 on the Claude API and on Amazon Bedrock, Google Cloud and Microsoft Foundry. The models overview gives a 1,000,000-token context window, 128,000 output tokens on the synchronous endpoint (300,000 on the batch endpoint behind a beta header), a June 2026 training and reliable-knowledge cutoff, adaptive thinking that is always on, and a default effort of medium. The price is $4 per million input tokens and $20 per million output tokens, with cache reads at 5 per cent of the input price and cache writes at $5; a faster serving tier is sold at $8 and $40. What is unusual, and worth more to a buyer than any benchmark, is the lifecycle: Anthropic’s deprecations page gives this model a retirement floor of not sooner than 22 September 2027 on Anthropic-operated platforms, alongside a commitment of at least 60 days’ notice before any publicly released model is retired — and it does the same for every other model in the table, including the ones already retired. Amazon Bedrock and Google Cloud set their own dates, which is the detail to check if you buy through a cloud marketplace rather than direct. Artificial Analysis, measuring independently, scores it 58 on version 4.3.2 of its Intelligence Index and puts its blended rate at $2.94 per million tokens on a 7:2:1 cache-hit, input, output mix; the same page records 260 million output tokens to complete the index against a median of 88 million, so the price per token and the cost per task are not the same question. Anthropic’s own figures — 66.4 per cent on Terminal-Bench 4.0, 54.4 per cent on FrontierCode v1.1, 57.8 per cent on CursorBench 4.0, 40.0 per cent on AutomationBench and 81.8 per cent partial on OSWorld 2.0 — are its own evaluation and are not stored as scores. The temperature, top_p and top_k parameters are not accepted on this generation and return a 400 error if set.