The coding jump, in numbers
Opus 5 is, by how the benchmarks tell it, the strongest flagship model on agentic coding. On SWE-bench Pro it scores 79.2% against Opus 4.8's 69.2%: ten points. For scale, Sonnet 5 on the same test sits at 63.2%, while the frontier models Fable 5 and Mythos 5 sit around 80%. On SWE-bench Verified the reported score is even higher.
A methodology note, because the numbers only count if they're honest: apart from SWE-bench Verified, these figures come from third-party reports in the first days after launch. The order of magnitude is solid, the individual percentages should be taken as indicative until the full official evals are out. If you need the non-technical picture, start from what Opus 5 is.
CursorBench and cost per task
The most interesting figure for developers isn't the quality peak, it's the quality/cost ratio. On CursorBench 3.2, at max effort, Opus 5 gets within about half a point of Fable 5 while paying half the cost per task. Anthropic also claims that at the high, xhigh and max effort levels, Opus 5 delivers more than any other model at the same cost.
Translated for a dev team: you can keep a flagship model on the whole flow — from reading the code to refactoring to debugging — without the electricity bill forcing you to downgrade on the heavy tasks. This is where the new model changes habits, not just benchmarks.
The effort dial for programmers
The effort dial is especially useful in development, where tasks differ a lot in difficulty. Generating a trivial unit test and designing a complex schema migration don't deserve the same reasoning budget.
With Opus 5 you raise the effort on architectural tasks and lower it on mechanical work. The result is a more predictable cost per session and, on hard tasks, more verification iterations before declaring done — Anthropic points precisely to a model "better at checking its own work and iterating until it succeeds." For anyone running agents autonomously, it's the difference between an agent that delivers and one that stops halfway.
Want to put Opus 5 to work on your code?
30 minutes to discuss your specific case.
The two API betas that matter
Along with Opus 5, Anthropic shipped two beta features aimed at anyone building agents on the API.
The first: changing tools mid-conversation without invalidating the prompt cache. An agent can gain or lose a tool on the fly — for example unlocking database access only after a check — without paying for the whole context again. On long flows it's a real saving. The second: automatic fallbacks. If a request is blocked by the safety classifiers, it gets routed to another available model instead of failing. Fewer unhandled errors in production. These are the finishing touches that separate a prototype from an agent that holds up under real traffic. For the foundations, see the guide on Agent SDK and autonomous agents.
What changes in Claude Code
In Claude Code, Anthropic's terminal tool, Opus 5 becomes the recommended engine for serious work: more reliable on multi-file tasks, better at verifying and correcting before wrapping up, with effort you can raise when you need depth. Fast mode, faster at double the price, is handy when latency bothers you more than cost — typical of dense interactive sessions.
For a team already working in Claude Code, the upgrade doesn't require rewriting anything: change the model underneath and re-measure. If you're bringing Claude Code or agents into a business context, with control over permissions and security, look at Claude Code for business.
Is the upgrade worth it?
For anyone developing with Claude, yes, almost always. Same list price as Opus 4.8, clearly better coding performance, cost control with the effort dial and two betas that remove friction from agents. The one caveat: re-run your own tests and evals, because a new model changes the outputs and what was tuned before needs re-verifying.
At Maverick AI we build agents and integrations with Claude Code, Agent SDK and MCP, and bring them into production with the controls that are needed. If you want to put Opus 5 to work on your code or your agents, get in touch.