Anthropic released Claude Opus 5 on 24 July 2026, two months after Opus 4.8. The line most people took away from the launch is that frontier intelligence now costs half of what it did. That is true in one specific sense and misleading in another, and the difference is worth about a hundred percent on your budget if you get it wrong.

Here is what actually shipped, what the numbers say, and what a small business should do about it.

What Anthropic actually shipped

Opus 5 is priced at $5 per million input tokens and $25 per million output tokens. It carries a 1M-token context window, which is both the default and the maximum, with 128K maximum output tokens. According to Anthropic’s own documentation, thinking is now on by default: the model decides when and how much to reason on each turn, and an effort parameter running from low through medium, high, xhigh and max is the control for how deep that goes.

It is available through the Claude API, Amazon Bedrock, Google Cloud and Microsoft Foundry. On the consumer side it became the default on Claude Max and the top model available on Claude Pro. Opus 4.8 stays available everywhere it was before.

Two smaller changes matter more than they look. The minimum cacheable prompt dropped from 1,024 tokens to 512, so short prompts that could never be cached before can be cached now with no code change. And batch processing halves both rates, to $2.50 and $12.50 per million tokens, for work that does not need an immediate answer.

The price story is being misread

This is the part worth slowing down on.

Opus 5 costs $5 per million input tokens. Opus 4.8 also cost $5 per million input tokens. Anthropic’s documentation states the pricing is unchanged from the previous Opus model. Nothing about this release made the Opus line cheaper than it was in May.

The “half price” framing is a comparison against a different model. Claude Fable 5, which launched in June, is priced at $10 per million input tokens and $50 per million output. So Opus 5 delivers something close to Fable-class capability at half of Fable’s rate. That is a real and significant claim. It is not a price cut, and if you budgeted on the assumption that your existing Opus spending is about to halve, it will not.

The practical version: your bill only falls if you were previously paying Fable rates for work that Opus 5 can now handle, or if you start using batch processing and lower effort levels. The headline number is not where the savings live.

What the benchmarks say, and what they do not

The reported results are strong. On OSWorld 2.0, a computer-use benchmark, Opus 5 scored 70.6%. On Frontier-Bench it reached 43.3%, more than double the Opus 4.8 result. On ARC-AGI 3, a test aimed at novel reasoning rather than memorised patterns, it moved from 1.5% to 30.2%.

That ARC-AGI jump is the one worth noticing, because it measures the thing benchmarks usually fail to measure: performance on problems the model has not effectively seen before. A twentyfold move on that axis says more than another point or two on a coding leaderboard.

Two caveats, though, and they apply to every model launch regardless of vendor. Benchmark figures at release come from the company shipping the model or from early third-party runs, under conditions chosen by whoever ran them. And a benchmark score is not a promise about your workload. A model that tops an agentic coding test will not necessarily write better product descriptions for your shop. The only number that matters for your business is the one you measure on your own tasks.

What changes for a small business

For most small and mid-sized companies, none of this requires action this week. The models available in your existing tools got better and the price of those tools did not move. That is the whole story for a typical user.

Where it does change something is anywhere you were priced out. Three concrete cases:

Long-document work. A 1M-token context window means an entire contract set, a full site export or a year of support tickets can go into a single request rather than being chopped into pieces and stitched back together. Chunking was always where quality quietly leaked.

Cost control that is now a dial, not a guess. The effort parameter lets you spend fewer tokens on routine work and more on the hard cases, in the same model. Combined with batch processing at half rate for anything that can wait, that is a bigger lever on a monthly bill than any headline price.

Verification-heavy tasks. Anthropic notes the model now verifies its own work without being asked, to the point that older prompts telling it to “add a verification step” cause it to over-verify. If you have prompts carried over from earlier models, they may now be working against you.

None of that needs a migration project. It needs someone to look at what you are already paying for and whether the settings still match the model.

Why this matters for how people find you

There is a second-order effect that has nothing to do with your API bill.

Every time frontier capability gets cheaper relative to what it replaces, more of it ends up inside consumer products: assistants that browse, summarise and answer directly instead of handing over a list of links. That trend does not depend on any single release. It has been the direction of travel for two years, and each launch of this kind moves it along.

The consequence for a business is unglamorous and consistent. A growing share of the people forming an opinion about you will do it through a machine-generated answer they never trace back to your site. Whether your business appears in that answer, and whether what it says is accurate, is not something you control through advertising. It comes down to whether the facts about you are clear, consistent and available in places these systems actually read.

That is a slower problem than a model launch, and a more durable one. The pricing on this release will be old news within a quarter. The question of whether an assistant describes your business correctly will not be.