Anthropic released Claude Opus 5 on 24 July 2026 at $5 per million input tokens and $25 per million output tokens, the same price as the Opus model it replaces. For a UK small business the useful part is not the benchmark table. It is the effort setting: a dial that lets you buy deep reasoning only on the jobs that need it, and pay considerably less on the ones that do not.
What is Claude Opus 5?
Claude Opus 5 is Anthropic's current model for complex coding and business work, released on 24 July 2026. Anthropic describes it as coming close to the frontier intelligence of Claude Fable 5 at half the price. It went live the same day on the Claude API, claude.ai, Claude Code and Claude Cowork, it is now the default model on the Max plan, and it is the strongest model available to anyone on Claude Pro.
The specifications that matter day to day are in Anthropic's models overview: a context window of 1 million tokens, up to 128,000 output tokens in a single response, and a reliable knowledge cutoff of May 2026. Anthropic's own rough conversion puts 1 million tokens at about 555,000 words. In plain terms, a year of contracts, policies or board packs fits into one conversation without anyone having to chop it up first.
What does Claude Opus 5 cost?
The API price is $5 per million input tokens and $25 per million output tokens, unchanged from Opus 4.8, 4.7, 4.6 and 4.5. If you use Claude through a subscription instead, Claude Pro is $17 per month billed annually or $20 billed monthly, and a standard Claude Team seat is $20 per user per month billed annually or $25 monthly. All list prices, before tax.
Two discounts change the arithmetic of anything you automate, and both are documented on Anthropic's pricing page. The Batch API takes 50% off input and output alike for work that does not need an answer this second, which puts Opus 5 at $2.50 and $12.50 per million. Prompt caching charges cache hits at 0.1 times the base input price, so a long-standing instruction set or reference document costs a tenth as much to re-send on later requests as it did the first time.
There is also a fast mode in research preview for Opus 5, running roughly 2.5 times faster at double the price: $10 and $50 per million. That buys latency, not capability.
What is the effort setting, and why does it matter for a small budget?
Effort is a request-level parameter that tells Claude how many tokens to spend. Opus 5 supports five levels, low, medium, high, xhigh and max, with high as the default. Anthropic's documentation is explicit that it affects every token in the response, not just reasoning: text, tool calls and thinking alike. Lower effort means fewer tool calls and less preamble.
That is the line that matters to a bill. Most of the money in a working AI system is not spent on hard questions. It goes on volume: sorting an inbox, pulling fields off invoices, tagging enquiries. Anthropic's own guidance for Opus 5 is to start at the default and then use low and medium liberally as your primary control for token cost and response time wherever your evals show quality holds.
Two caveats before anyone reaches for the dial. Effort is described as a behavioural signal rather than a hard token budget, so at lower levels Claude will still think on genuinely difficult problems, just less than it would at higher ones. And changing effort between requests invalidates prompt caching, so pick a level per workload and hold it steady rather than varying it inside one conversation.
What can a small business actually do with it?
Three patterns are worth a look, and they follow from price and context window rather than from cleverness. The first is whole-corpus questions. With a 1 million token window you can put an entire handbook, contract set or year of minutes in front of the model at once and ask across all of it, rather than building retrieval plumbing to feed it a paragraph at a time.
The second is high-volume, low-difficulty work run on the cheapest model that holds up. Anthropic's pricing documentation works a public example: 10,000 support conversations at roughly 3,700 tokens each, run through Claude Haiku 4.5, come to about $37 in total. The figure itself is not the point. The point is that model tier and effort level, not the number of tasks, are what set the bill.
The third is trying agentic work before commissioning any of it. Claude Code and Claude Cowork are included in Claude Pro at $17 per month billed annually, which is a modest price for finding out whether an assistant that reads your files and drafts the work earns its place in your business.
What this release does not change
The benchmark numbers in the launch post are vendor-reported, measured on tests Anthropic chose, and none of them describes your workload. The effort setting is not new either: the five levels have been in place since Opus 4.7, and effort as a parameter goes back to Opus 4.5. Anthropic itself advises running a fresh effort sweep on your own evaluations rather than carrying settings over from an earlier model, which is a fair admission that the right level is specific to the job and cannot be read off a table.
The pattern that works is unglamorous. Pick one task you already do repeatedly. Write down what a good output looks like. Run it at two effort levels and compare the results against that written standard. A model launch changes what the ceiling costs; it does not tell you which of your own jobs is worth automating, and nothing in this one changes that. If you want a structured way through that question, an AI Audit is where we normally start.