Claude Opus 5 Puts Near Frontier Agents at Half the Price. Here Is the Math for Your Business.

July 27, 2026. Anthropic released Claude Opus 5 on Friday, and the number that matters is not the one in the price list. The price did not move at all: Opus 5 costs the same $5 per million input tokens and $25 per million output tokens as Opus 4.8 did. What moved is how much work you get for that money. On Anthropic's own evaluations the model reaches near the frontier intelligence of Claude Fable 5 at half the price, and finishes more tasks per dollar than anything else the company has shipped. If you run AI automation for a small business or an agency, that is a reason to reopen the spreadsheet on every project you shelved because the model was too expensive or too unreliable to trust unattended.

What Anthropic actually shipped

1. Claude Opus 5 went live on July 24 across the Claude API, claude.ai, Claude Code and Cowork, priced at $5 per million input tokens and $25 per million output tokens, identical to Opus 4.8. Anthropic states it comes close to the frontier intelligence of Fable 5 at half the price. The launch post is the primary source for every figure below.
2. It is now the default model on Claude Max and the strongest model available on Claude Pro. A Fast mode runs around 2.5 times the default speed at twice the base price.
3. An effort setting lets you trade intelligence against tokens per request. Anthropic publishes performance by effort level, and says Opus 5 beats every other model at a given cost on high, xhigh and max effort on CursorBench.
4. On Frontier-Bench v0.1, Anthropic reports Opus 5 surpasses all other models and more than doubles Opus 4.8's score at a lower cost per task. On CursorBench 3.2 at max effort it lands within 0.5% of Fable 5's peak score at half the cost per task.
5. On Zapier's AutomationBench, which measures whether a model can complete business tasks from start to finish, Anthropic reports a pass rate around 1.5 times the next best model at the same cost per task, and says even the lowest effort setting passes more tasks than any other model. On OSWorld 2.0, a computer use benchmark, it reports beating Fable 5's best result at just over a third of the cost.
6. Two developer changes shipped in beta alongside it. Mid-conversation tool changes let you swap which tools Claude can use without invalidating the prompt cache. Automatic fallbacks let API requests that trip a safety classifier route to another model instead of failing.

Why cost per task is the only number worth tracking

Almost every AI budget conversation we have with founders is priced in dollars per million tokens, and almost every one of them is wrong. Tokens are an input. What you actually buy is a finished task: a proposal drafted, a lead enriched, a support thread resolved, a report reconciled. A cheaper model that needs four attempts and a human rescue is more expensive than a pricier one that lands on the first try.

That is what changed on Friday. The token price is flat, so any improvement in how few turns and tokens a task takes is a direct cut to your cost per completed task. Anthropic's own framing leans on this: its benchmark charts plot performance against cost per task rather than raw scores. Early access customers report the same shape. The legal AI company Harvey said Opus 5 held quality while generating 26% fewer tokens on average than Opus 4.8 at max reasoning. Box reported an 8% overall gain over Opus 4.8, with 11% on data analysis and 17% on due diligence workflows. Those are customer claims published by the vendor, not independent tests, so treat them as directional.

What it means for operators

Three practical shifts. First, the projects you killed on ROI need re-scoring, not rebuilding. Most automation ideas that failed a business case in the last year failed on cost per successful run or on supervision overhead, and both just improved without you changing a line of your workflow. Second, the effort setting turns model choice into a dial rather than a switch, which means one workflow can run cheap by default and escalate only where quality actually breaks. Third, the automatic fallback beta quietly removes a real production failure: an agent that stops mid run because a classifier flagged a request. If you have ever had a client automation die at 2am for a reason nobody could explain, that is the class of failure this addresses.

There is also a supervision angle. Anthropic reports Opus 5 scored 2.3 on overall misaligned behaviour in its automated behavioural audit, the lowest of its recent models, with the lowest rates of deceptive behaviour and the best record at avoiding reckless actions with hard to reverse side effects. That does not mean you remove the human. It means the failure rate you are budgeting supervision against is probably lower than the one you measured six months ago, and you should measure it again rather than assume.

How to use this in the next week

1. Pick one shelved workflow and re-run its business case in cost per completed task. Count retries, human corrections and rework, not tokens.
2. Set effort per workflow rather than globally. Start at a low setting, escalate only the steps where output quality actually fails a check.
3. Turn on automatic fallbacks before you need them, so a flagged request degrades to a working model instead of breaking the run.
4. If your agents change tools mid run, audit where you were paying to rebuild context. The mid-conversation tool change beta removes that cost.
5. Re-test before you switch anything client facing. Run 20 real tasks from your own backlog and compare cost, turns and corrections. Behaviour changed, so your old prompts are not guaranteed to be optimal.

The honest caveats

Every benchmark figure above is Anthropic's own. Its Frontier-Bench footnote says the results come from an internal run on the mini-SWE-agent harness, averaged over five attempts per task, with Opus 4.8 serving as fallback on safety classifier refusals. Independent write ups published since Friday quote conflicting Frontier-Bench numbers for the same model, which is a good reason to cite the vendor's relative claims and run your own test rather than repeating a specific score. Trade coverage including VentureBeat frames it the same way we do, as a cheaper path to agentic and enterprise workflows rather than a raw capability jump. And Anthropic is explicit that Opus 5 remains behind Mythos 5 on biology and offensive cybersecurity work, with cyber classifiers it expects to intervene around 85% less often than Fable 5's.

The strategic point stands regardless of which benchmark you believe. The cost of running a competent agent fell again this month, for the third time this quarter, and the businesses that win are the ones re-testing their assumptions on that cadence instead of once a year. If you want that testing done properly, our team does exactly this work through our AI automation agency, or you can hire an AI engineer to run the evaluation inside your own stack.