Meet the New Claude Opus 5: Frontier-Class Agentic Coding and Computer Use at Unchanged Opus Pricing

Anthropic quietly rolled out Claude Opus 5 today, and the headline number is the one that didn't change: $5 per million input tokens, $25 per million output. Same price tag as Opus 4.8, the model it replaces. Anthropic's pitch is that Opus 5 gets close to Claude Fable 5's intelligence for half the cost, which is a bold claim to make about your own top-tier product, but it's now the default on Claude Max and the strongest option on Claude Pro, so the company is putting its money where its mouth is.

The more interesting story is buried in the API details, not the marketing line. Thinking is on by default now. Previously, with Opus 4.8, a request ran without any extended reasoning unless you explicitly flipped on thinking mode. Opus 5 flips that assumption. Every request thinks unless you tell it otherwise, and the effort parameter now governs how deep that thinking goes. That sounds like a minor knob change until you remember that max_tokens covers both the thinking budget and the actual response text, which means anyone with hardcoded token limits from the Opus 4.8 era needs to go back and check their math.

There's an actual breaking change hiding in here too. Try to disable thinking while effort is set to xhigh or max, and the API now throws a 400 error. Anthropic enforces this per request, so developers either cap effort at high or strip the thinking field entirely. It's a small technical detail, but it's the kind of thing that silently breaks production pipelines at 2 a.m. if nobody reads the changelog.

Anthropic is also telling people to rip out their old prompting habits. Verification instructions that used to be standard practice, things like asking the model to add a final check step, now cause Opus 5 to over-verify, because the model already does that work internally. That's a subtle admission that the underlying reasoning behavior has shifted enough to make old prompt engineering counterproductive, not just unnecessary.

On the specs side, Opus 5 keeps context at a flat 1 million tokens, no smaller tier to choose from. Output tops out at 128k tokens through the standard Messages API, but the Message Batches API can push to 300k with a beta header. Anthropic also cut the minimum cacheable prompt size in half, from 1,024 tokens down to 512, which should help smaller, more frequent requests benefit from caching discounts that used to be out of reach.