Claude Opus 5 does the hard technical work without flagship pricing

For most of this year, businesses doing anything technical with AI have faced an annoying choice. The model that could actually handle the work, Claude Fable 5, ran at $10 per million input tokens and $50 per million output. Everything cheaper was noticeably worse at the jobs that matter: reasoning across a large codebase, extracting structure from messy documents, running a multi-step process without losing the thread. So you either paid flagship rates or you accepted a model that needed more supervision.

Anthropic released Claude Opus 5 on 24 July at $5 per million input tokens and $25 per million output. That’s the same price as Opus 4.8, the model it replaces, and half what Fable 5 costs. On Anthropic’s published benchmarks it matches or beats Fable 5 on most technical work.

If you already have Opus 4.8 running in a production workflow, switching costs you a config change and nothing else.

The numbers, and how much to trust them

On Frontier-Bench v0.1, which tests agentic coding in a terminal, Opus 5 scores 43.3% against Fable 5’s 33.7% and Opus 4.8’s 21.1%. On CursorBench at maximum effort it lands within 0.5% of Fable 5’s best score at half the cost per task, and half a percent is inside normal run-to-run variation. On OSWorld 2.0, which measures a model driving a computer, it hits 70.6% against Fable’s 66.1%, and it beats Fable’s best result at roughly a third of the cost.

Two caveats before you act on any of that. These are the vendor’s own numbers, published to sell the vendor’s own model, and vendor benchmarks are chosen the way a real estate photo is framed. They’re also aggregate scores on generic tasks, and your work isn’t generic. A model that tops a coding benchmark can still be mediocre at reading your particular supplier invoices.

The useful move is boring: keep twenty or thirty of your actual tasks with known-correct answers, run them against any model you’re considering, and score them yourself. That takes an afternoon and tells you more than every launch chart put together. We go through this properly in getting from proof of concept to production, because the gap between a good demo and a dependable process is where most AI projects fall over.

The effort setting is where the money actually is

Opus 5 has five effort levels: low, medium, high, xhigh and max. They control how much the model thinks before answering, and the difference in cost is substantial. On CursorBench the ladder runs from about $2.50 a task at low effort to about $8 at max.

Opus 5 at low effort scores 62.8% on that benchmark, better than Opus 4.8 managed at maximum effort, for less than half the money. Most business workloads are not hard problems. Classifying an email, pulling fields out of a PDF, drafting a first response, checking a form against a rule set: none of that needs the model reasoning at full depth, and running everything at maximum is how you turn a sensible AI budget into a surprising one.

Pick the effort level per job. One system-wide setting will be wrong for most of what passes through it. Save the expensive levels for long-running agentic tasks, code changes across multiple files, and anything where a wrong answer costs more than the tokens do.

There’s also a fast mode at 2.5 times the output speed for twice the base price. That puts you back at Fable rates, and what it buys is throughput. The answers are the same ones you’d have got anyway. Worth it when a customer is sitting there waiting, hard to justify for a nightly batch job.

Where the flagship still makes sense

Opus 5 doesn’t win everywhere. Fable 5 stays ahead on legal agent tasks, 13.3% to 11.7%. Mythos 5 is well clear on professional health questions at 66.0% against 59.8%. GPT-5.6 Sol takes the DeepSWE coding benchmark at 72.7%, with Opus 5 on 68.8%. On FrontierCode the difference between Opus 5 and Fable 5 is a tenth of a percent, which is a tie by any honest reading.

So if your work is legal analysis or clinical content, the flagship premium is still buying you something measurable. For software engineering, workflow automation, document processing and agentic operations, it isn’t, and paying double for it is a habit nobody has gone back and questioned.

Price also isn’t the only question worth asking. Where your data goes still matters more than which model is fastest this quarter, particularly if you’re handling client files under privilege or privacy obligations. We’ve written about what private AI actually means and data sovereignty for Australian businesses, and neither of those calculations changes because a cheaper model turned up.

Two things that will catch you on the way in

If you’re swapping the model in an existing integration, there are two behaviour changes that cost people an afternoon each.

Thinking is now on by default. Previously a request that didn’t ask for extended reasoning didn’t get any. Now it does, which means requests cost a little more than you’d modelled, and any code that sized its token limit tightly around the expected answer will start truncating mid-response. Check your limits before you switch, not after a customer reports a half-finished reply.

Second, you can only turn thinking off at high effort or below. Ask for thinking disabled at xhigh or max and the API returns an error rather than quietly doing something reasonable. It’s a one-line fix once you know, and a confusing hour if you don’t.

The pattern underneath the launch

Anthropic has now done this three times in a row. Opus 4.7, then 4.8, now Opus 5, all at $5 and $25, each one better than the last at the same rate.

That has a practical consequence for how you build. If what you can buy for a dollar keeps improving every few months, hard-wiring one model into your systems is the expensive choice. Keep the model name in a single configuration value you can change in a minute. Keep your prompts and your scoring set in version control. Build the workflow around the job being done, so swapping the model underneath means an afternoon and a re-run of your tests.

Anyone set up that way has been able to take each of these releases with a one-line change. Anyone who built around a single model’s quirks is now weighing a rewrite against staying on something slower and more expensive, and that decision gets harder with every release.

If you’re working out which model fits which part of your business, or you’ve got an AI process that’s costing more than it should and you can’t see where the money goes, get in touch and tell us what the process does and what it’s costing. The free AI readiness assessment is a quicker way to see where you stand if you’re earlier than that, and our applied AI work is where this usually starts: one process, measured properly, before anything gets rolled out further.