Claude Opus 5 Shipped. The Real Change Is the Effort Dial, Not the Benchmarks

Three weeks after I wrote about Sonnet 5 quietly becoming the Claude Code default, Anthropic did it again. Claude Opus 5 shipped July 24, 2026, it's now the default on Claude Max and the strongest model on Pro, and the model ID in my terminal changed a second time without me touching a config file.

I'll get to the benchmarks, because they're genuinely good. But if you're building anything on the API or Claude Code, the benchmarks aren't the thing that will bite you this week. A default-behavior change will.

What actually shipped

Opus 5 (claude-opus-5) replaces Opus 4.8 at the top of the Claude lineup: 1M-token context window (fixed, not a toggle), 128K max output, and pricing unchanged from Opus 4.8 at $5 per million input tokens and $25 per million output. Claude Fable 5, Anthropic's most capable model, runs $10/$50, exactly double. Opus 5 is Anthropic's attempt to close most of that intelligence gap without the price tag.

On coding-specific evals, Opus 5 is the new state of the art among widely available models: on Frontier-Bench it more than doubles Opus 4.8's score, and on CursorBench it lands within half a percentage point of Fable 5's peak result while costing half as much per task. Anthropic's own framing is blunt about it: Opus 5 delivers "frontier intelligence at half the cost of Claude Fable 5." Knowledge cutoff is May 2026, the most current of any Claude model right now, which matters more than it sounds for anything touching recent libraries or APIs.

None of that required me to change a single line of code. This did.

Thinking is on by default now. That's a bigger deal than it sounds

Every earlier Opus model ran without extended thinking unless you explicitly asked for it. Opus 5 flips that: it decides on its own, per turn, whether and how much to think, and the lever you actually control is effort, not a thinking toggle. The ladder is low, medium, high, xhigh, max, defaulting to high on the API and in Claude Code.

Two things fall out of that, and one of them is a real breaking change if you're hitting the API directly (Claude Code and claude.ai handle it for you):

You can no longer disable thinking at xhigh or max effort. Set thinking: {"type": "disabled"} at those levels and you get a 400 back. If you're running an Agent SDK build or a custom Messages API loop with thinking explicitly turned off and effort cranked up, that request breaks on upgrade, not gracefully, with an error. Check any code that touches both fields before you flip the model ID.

Old "please verify your own work" instructions now actively hurt you. Opus 5 verifies without being told to. Anthropic's own migration notes call this out directly: carry over a "double-check your answer" or "use a subagent to verify" instruction from an Opus 4.8 prompt and Opus 5 over-verifies, spending tokens re-checking work it already checked. I've been teaching cohort attendees to bolt a verification step onto every agentic prompt for the better part of a year. That advice just got an asterisk on it.

There are two smaller wins buried in the same release notes. The minimum cacheable prompt length dropped from 1,024 tokens to 512, so shorter prompts that couldn't hit prompt caching on Opus 4.8 now can, with no code changes. And mid-conversation tool changes are in beta: you can add or remove tools between turns of a session while keeping the prompt cache warm, instead of resending a fixed tool list for the life of a conversation. Useful if you're building an agent whose available tools should change as the task narrows.

What I'm actually doing differently

I audited every place I have thinking explicitly disabled. If any of those requests also set effort to xhigh or max, that's a live bug the moment I migrate the model ID, not a theoretical one. Took ten minutes to grep for it across my Agent SDK code and one MCP server.

I'm pulling verification instructions out of prompts, not adding more. For a year the standard advice, mine included, was "always tell the agent to check its own work." Opus 5 does that on its own by default. The new job is trimming boilerplate, not writing more of it, and watching whether that actually holds up on gnarlier multi-step builds.

I'm treating effort as the real dial now, not thinking-on-or-off. high is the default and it's a reasonable one, but Anthropic's own guidance is to start there and move in either direction based on your evals: step down to medium or low where quality holds and you want the tokens and latency back, step up to xhigh or max for the work that's actually hard. That's the same "model, effort, thinking" framing I already teach in the Claude Camp curriculum. Opus 5 just made the effort slider the one that matters most.

If you're on Claude Code, you're likely already on Opus 5 without having asked for it, same as Sonnet 5 three weeks before it. Go check what your requests do when thinking is off and effort is high before something quietly breaks in production.

Sources: Introducing Claude Opus 5 · What's new in Claude Opus 5 · Models overview · Meet the New Claude Opus 5