Claude Opus 5 Explained (2026): Benchmarks, Pricing & How It Compares

at a glance

For most of 2026, the top of Anthropic’s lineup meant paying flagship prices. Claude Opus 5, released on 24 July, changed that. Anthropic’s own line for it is blunt: a model that “comes close to the frontier intelligence of Claude Fable 5 at half the price.” After going through the launch data, that framing mostly holds up.

The important part isn’t that Opus jumped from 4.8 to 5. It’s that the jump came at the same price the old Opus charged. You get a materially stronger model for what you were already paying.

A note on how we cover this. Opus 5 is a day old as we publish. We haven’t run it through our own build brief the way we test app builders, so this is research-grounded analysis, not a hands-on score. Every number below is Anthropic’s own, and it’s dated and linked.

What actually changed

Three things make Opus 5 worth paying attention to.

The price-to-capability reset. Opus 5 costs $5 per million input tokens and $25 per million output. That’s identical to Opus 4.8 and half of Fable 5’s input rate. Anthropic reports it roughly doubles Opus 4.8’s score on Frontier-Bench, its coding benchmark, while charging the same.

The effort toggle. Opus 5 exposes a cost-versus-capability dial — low, medium or high effort per task. Turn it up for hard problems and deep reasoning; turn it down when you want a fast, cheap answer. On the API and in Claude Code it defaults to high. This is the feature the press latched onto, because it hands the cost decision to the user instead of baking it in.

Developer plumbing. Two beta additions: mid-conversation tool changes, so an agent can gain or lose tools without starting a new session, and automatic fallback routing, which lets the API retry on another model when a request needs it. Opus 5 also drops the data-retention requirement that Fable 5 carries, which matters for teams with strict data policies.

The benchmarks, in plain terms

Anthropic published a spread of results at launch. Here’s what they say, translated out of benchmark-speak.

On coding and software work, Opus 5 tops Anthropic’s Frontier-Bench and comes within roughly half a percent of Fable 5 on CursorBench while costing half as much per task. On general knowledge work it scores about three times higher than the next-best model on ARC-AGI-3, and clears Fable 5’s result on the OSWorld computer-use benchmark at a third of the cost. On science, Anthropic measured gains of 10.2 percentage points on organic chemistry and 7.7 on protein prediction over Opus 4.8.

There’s a real limit worth naming. Opus 5 sits behind Anthropic’s restricted Mythos 5 model on biology research and cyber-exploitation tasks, and its safety classifiers are tuned to be about 85% less restrictive than Fable 5’s on narrow cyber work. For ordinary building, none of that touches you. For frontier security or bio research, it does.

Opus 5 vs Opus 4.8 vs Fable 5

The takeaway is simple. Against 4.8, there’s no contest — same price, much stronger, so migrate. Against Fable 5, it’s a value call: Opus 5 gives you most of the flagship for half the input cost, and the gap only matters at the extreme end of agentic work.

Who should use Claude Opus 5

How to build with Opus 5 today

You don’t need to write API code to use it. Opus 5 is live on the Claude API, in Claude Code, and across the chat apps — it’s the default on Claude Max and included with Claude Pro. But if your goal is to ship software, the fastest path is an AI app builder that runs on Claude and points at a current model for you.

Cursor — whose own CursorBench is one of the benchmarks Opus 5 nearly tops — lets you select Claude as the model behind its agent. Prompt-to-app builders like Lovable turn a plain-English brief into a working, deployed app on top of Anthropic’s models, with no setup at all.

The bottom line

Claude Opus 5 is the model most builders should be using right now. It matches the old Opus on price, roughly doubles it on coding, and gets close enough to the Fable 5 flagship that paying double is hard to justify outside the most demanding agent work. If you’re on Opus 4.8, move. If you’re deciding where to start, start here — and if you’re deciding how to put it to work, do it inside a tested app builder rather than from scratch.

Frequently asked questions

Keep reading