Update: Claude Opus 5 ships at 4.8 pricing, with an effort dial
Pricing holds at 4.8's; the wins are a 1M context window, thinking now on by default, near-Fable-5 scores, a 2.5x Fast mode, and day-one Copilot support.
Copy markdown
Live now, and the price didn't move
claude-opus-5 is available today on the Claude API, Bedrock, Vertex, and Foundry at $5/$25 per million tokens — identical to Opus 4.8 — with a 1M-token context window and 128k max output. For most calls, swapping the model string is the whole migration.
One effort dial, from low to max
A new effort setting — low, medium, high, xhigh, max — lets you trade intelligence for tokens per request instead of switching models. Anthropic pitches it as the fix for runaway bills: dial down for cheap bulk work, up for hard reasoning.
The migration trap: thinking is on by default
Unlike 4.8, Opus 5 thinks by default, and you can only disable it at effort high or below — sending thinking:{type:'disabled'} at xhigh or max returns a 400. Audit any pipeline that assumed thinking was off.
Near-Fable-5, at half the cost
Anthropic's numbers: Opus 5 more than doubles 4.8 on Frontier-Bench, lands within 0.5% of Fable 5 on CursorBench 3.2 at half the cost per task, and triples the next model on ARC-AGI 3. It's now the default on Claude Max and the top pick on Pro.
A 2.5x Fast mode, if you'll pay double
An API-only research-preview Fast mode runs about 2.5x faster for exactly double the price ($10/$50 per million). It's not on Bedrock, Vertex, or Foundry yet — aimed at latency-bound agent loops where wall-clock beats token cost.
Already in GitHub Copilot
Opus 5 is in the Copilot model picker today for Pro+, Max, Business, and Enterprise across VS Code, JetBrains, Xcode, the CLI, and the cloud agent, billed at provider list price. Enterprise admins must flip the Opus 5 policy on first.
Swap tools mid-conversation, keep your cache
Two platform betas ship alongside: you can change the available tools mid-conversation without invalidating cached prompts, and safety-flagged requests now auto-fall-back to another model instead of erroring out.