Anthropic launched Claude Opus 5 on July 24 as a generally available model on all its platforms, under the API identifier claude-opus-5. It costs $5 per million input tokens and $25 per million output, the same rate card as Opus 4.8, and Anthropic pitches it as close to Claude Fable 5 at half the price. That half is arithmetic: the pricing documentation lists the Fable tier at $10 and $50.
Where it runs and who gets it
Opus 5 is available on the Claude API, Amazon Bedrock, Google Cloud and Microsoft Foundry, with a 1.1 times multiplier for US-only inference. In the Claude apps it is included in Pro, Max, Team and Enterprise, not in Free. At launch Anthropic called it the new default model on Max and the strongest model on Pro; whether it stayed the default on Max after Fable 5.1 arrived on September 1 is not stated on any page.
It keeps a context window of one million tokens at standard pricing, and effort is the main control, from low to max, with high as the default.
What the bill depends on
The pricing documentation sets five-minute cache writes at $6.25 per million tokens, one-hour writes at $10 and cache reads at $0.50, and the Batch API halves the base rates to $2.50 and $12.50. It also warns that Claude models from 4.7 onwards use a tokenizer that produces roughly 30% more tokens for the same text, which raises the effective cost of the same work.
Fast mode, which Anthropic says delivers up to 2.5 times the output speed, costs exactly twice the base price, $10 and $50, and remains a research preview rather than a general feature. It is supported only on Opus 5 and Opus 4.8, and not on Claude Platform on AWS.
General, preview or beta
The model itself, effort, the one-million-token context, caching and batch are generally available. Fast mode is a research preview. The two API features that shipped with the model are betas behind dated headers:
- Mid-conversation tool changes let a developer offer or withdraw tools between turns without invalidating the prompt cache.
- Automatic fallbacks route requests flagged by safety classifiers to another model, server-side or through SDK middleware. They do not work with the Message Batches API or on Bedrock, Google Cloud and Microsoft Foundry.
Safeguards and whose claims
Anthropic says Opus 5 was deliberately not trained on cyber tasks, and its cyber safeguards block binary-based scanning, penetration testing and exploit generation while leaving source-code vulnerability finding open, with its Cyber Verification Program as the route to fewer restrictions. It expects those classifiers to intervene about 85% less often than Fable 5’s. In chat, flagged requests fall back to Opus 4.8 by default, and Opus 4.8 stays on sale.
The leadership claims are Anthropic’s: state of the art on Frontier-Bench and GDPval-AA, near Fable 5 on CursorBench, “most aligned model to date”. The independent GDPval-AA leaderboard from Artificial Analysis places Opus 5 near the top, behind Anthropic’s own Fable 5.1.
The announcement says nothing about marking its output. Anthropic later tied its text watermark to models released after August 2, 2026, a date Opus 5 predates, although Anthropic has not said so about Opus 5 specifically.
Related reading
- Claude Opus 5 system card — Anthropic
- Fast mode — Claude documentation
- Fable 5 is back with new safeguards — IA al Día