Claude Opus 5 ships with a 1-million-token context window as both default and maximum, 128,000 output tokens in the standard API, and unchanged Opus pricing of $5 per million input tokens and $25 per million output. [1]
The paper’s Sunday brief on Opus 5’s window treated the number as a race. Monday’s specification is the role: Opus 5 replaces Opus 4.8 as the workhorse. Fable 5 remains the frontier flagship. Mythos 5 stays ahead on cybersecurity. [1]
Anthropic launched the model on July 24. It is the new default on Claude Max and the strongest model on Claude Pro. Fast mode runs about 2.5 times quicker at twice the price. Effort settings let a customer trade tokens for depth. On some high settings, thinking cannot be switched off. [1]
The company leans on its own suite: Frontier-Bench, CursorBench, ARC-AGI 3, GDPval-AA, OSWorld. It says Opus 5 more than doubles Opus 4.8 on Frontier-Bench at a lower cost per task, and that ARC-AGI 3 scores sit about three times the next listed model. Independent testers at Artificial Analysis ranked it first on intelligence and agentic indexes at launch. [1][2]
A million tokens is enough, in the marketing sentence, for a large codebase or a legal discovery set in one prompt. The production question is whether instruction-following holds at the far end of that window. Anthropic says it does. Customers will find out on invoices.
X still posts the 1M figure as if Anthropic had just crowned a new emperor. The launch page is more modest, and more interesting: a mid-tier model that can hold a book of evidence without charging like a myth.
Early-access notes are the usual parade. Cognition said Opus 5 approached Fable-level work on FrontierCode at half the cost. Cursor’s Sualeh Asif said it sat just under Fable 5 on CursorBench. Zapier said it completed a churn-prevention sequence no prior model passed. Those are vendor-selected stories. They are still more useful than a context-window meme. [1]
Opus 4.1 is being retired. Automatic fallbacks will send safety-flagged Opus 5 calls to another model instead of a hard block. Mid-conversation tool changes will no longer bust the prompt cache. That is the production tier talking: fewer broken sessions, same list price, a window wide enough to embarrass a filing cabinet. [1][3] The race is not over. The product, for once, is. MarkTechPost filed the launch as frontier-class coding at unchanged Opus rates. That hed is closer to Anthropic’s own hierarchy than the 1M-token memes. [3] A window is a capability. A price is a policy. Opus 5 is both, and still not the top of the catalog. Teams that paste an entire repository into a prompt will learn, this week, whether the model still knows which file matters. That is the only test that counts.
-- KENJI NAKAMURA, Tokyo