Anthropic shipped Claude Opus 5 on July 24, and the headline isn’t the intelligence. It’s the price. Opus 5 lands at roughly the capability level of Fable 5, Anthropic’s frontier model, at half the input price: $5 per million input tokens and $25 per million output, the same rate the previous Opus charged. If you’re a business weighing an AI project, that pricing move matters more to you than any benchmark chart.
What actually shipped
The checkable specifics: Opus 5 runs at $5/$25 per million tokens (unchanged from Opus 4.8), reaches near-Fable-5 performance on agentic coding benchmarks at half the cost per task, carries a 1M-token context window, and adds two controls worth knowing about. A fast mode runs about 2.5 times quicker at double the price, and an effort toggle (low, medium, high) lets each individual request trade cost against capability. It’s the new default on Claude Max plans and available in the API as claude-opus-5.
The usual caveat applies: launch-day benchmarks are vendor-flavored, and every model release since 2023 has claimed a breakthrough. The pricing is a fact, though, and it continues the only trend that has held steady through all of it: the cost of a unit of AI capability keeps falling, fast.
Why the price matters more than the benchmarks
We have our own data point on how cheap production AI has already become. CodeRaven, the AI code-review platform we built and operate, reviewed 1,164 pull requests and posted 14,552 findings over its first five months, for a total model bill of $161.20. We published that dataset in full. That bill was run on the previous generation of models. Releases like this one push the same work toward cheaper still.
The practical meaning for a business: model cost is now rarely the thing that decides whether an AI project pencils out. Engineering, evaluation, and integration are. Which is also why we stay deliberately unromantic about model choice; it’s an engineering decision made per task, not a loyalty program.
The effort toggle is the quiet story
The per-request effort setting deserves more attention than it’s getting. Production AI systems, especially agents, are mostly made of small, repetitive model calls with a few genuinely hard ones mixed in. Being able to run the routine calls cheap and reserve full effort for the decisions that matter is exactly how you engineer a system whose costs stay boring at scale. Our five-month bill looked the way it did because of scoping decisions like that, made before this toggle existed. Now the lever is built in.
What we’d actually do this week
If you already run AI in production and you have evaluation suites, testing Opus 5 against your real tasks is an afternoon of work, and the possible outcome is the same quality at half the spend. If you don’t have evals, that’s the finding: you can’t safely take advantage of releases like this one, and the next one, without them. Building that discipline in from the start is a core part of our custom LLM development work, and it’s the difference between a system that rides each price drop and one that’s welded to whatever model it launched on.
And if you’re still deciding whether AI fits your business at all, note what this release did to the downside risk: the cost of finding out keeps shrinking. That’s what our AI consulting assessments are scoped around.
Source: Anthropic’s announcement, with early benchmark detail from MarkTechPost.
Related reading: Claude Code vs Webflow vs WordPress — what the agentic coding stack means for how business websites get built.
