Anthropic released Claude Opus 5 on July 24, 2026, and the headline isn't intelligence. It's cost control. The model comes close to the performance of Claude Fable 5, Anthropic's most capable publicly available model, at half the price, and it ships with an effort setting that lets you decide how much compute each task deserves.

If you run AI-powered workflows in production, those two details matter more than any benchmark score. Here's a full breakdown of the pricing, how effort settings work, and what both mean for teams automating real work.

How much does Claude Opus 5 cost?

Claude Opus 5 is priced at $5 per million input tokens and $25 per million output tokens on the API. That's unchanged from its predecessor, Opus 4.8, and half the rate of Fable 5.

A few more pricing details worth knowing:

Opus 5 is the new default model on Claude Max and the strongest model available on Claude Pro, so subscription users get it without paying more. Developers can access it on the Claude API using the model string claude-opus-5. There's also a Fast mode that runs around 2.5 times the default speed at twice the base price, for latency-sensitive work.

The pricing itself isn't news. Getting substantially more capability at the same price is. But the bigger story for anyone watching their AI spend is the new control that ships with it.

What are effort settings in Claude Opus 5?

Effort settings are an adjustable dial that controls how much computing power the model devotes to a task. You can set it to low, medium, or high. At lower effort levels, Anthropic says the model preserves much of its performance while using fewer tokens, running faster, and costing less to operate.

Until now, cost control in AI orchestration meant model routing: send easy tasks to a cheap model and hard tasks to an expensive one. That works, but it forces you to maintain routing logic, test multiple models, and accept that the cheap model sometimes fumbles a task you thought was easy.

An effort dial moves that decision inside a single model. Data extraction from a clean invoice? Low effort. Reconciling conflicting records across three systems? High effort. One model, one integration, one set of behaviors to learn, with cost scaled to task difficulty instead of applied as a flat tax on everything.

And the low-effort mode isn't a token gesture. On Zapier's AutomationBench, which measures whether models can complete business tasks from start to finish, Opus 5's pass rate is around 1.5 times the next-best model at the same cost per task. Even at its lowest effort setting, it passes more tasks than any other model. The cheapest configuration of this model completes more end-to-end business tasks than competitors running at full strength.

Early access customers reported the same pattern. Harvey, a legal AI company, found Opus 5 at lower reasoning levels matched higher-effort performance while generating 26% fewer tokens than Opus 4.8 at max reasoning. A trading firm reported its strongest benchmark results using roughly a seventh of the reasoning tokens and under half the latency of the previous model.

Claude Opus 5 vs Fable 5: which one should you use?

Fable 5 remains Anthropic's smartest generally available model, and the company still recommends it for the most advanced projects, including autonomous work that runs for days. But the gap is narrower than the price difference suggests.

On CursorBench, a coding benchmark, Opus 5 at max effort performs within 0.5% of Fable 5's peak score at half the cost per task. On OSWorld 2.0, a computer use benchmark, Opus 5 surpasses Fable 5's best result at just over a third of the cost.

There's also a practical difference in how the two models spend. When Fable 5 launched, business customers criticized its burn rate, the high number of tokens it consumed completing tasks. Users blew through token budgets or ran up unexpected bills. Opus 5 was built as the everyday alternative: for most workflow automation, knowledge work, and agent tasks, it's the more economical default, with Fable 5 reserved for the genuinely hard problems.

The short version: default to Opus 5, escalate to Fable 5 when a task proves it needs the extra intelligence. That's a cheaper strategy than the reverse.

Why reliability is a cost line too

There's a second cost story in this release, and it's about failure rates.

Every failed automation run costs you twice; once for the tokens spent on the failure, and again for the human time spent catching it, diagnosing it, and rerunning it. In most production workflows, that second cost overshadows the first. Token bills are annoying. Silent failures that reach a customer are expensive.

Anthropic describes Opus 5 as a model that verifies its own work and recovers from errors without intervention. Customer reports back this up in concrete terms. One frontend testing company found the model opened its own pages in a browser at desktop and phone widths, caught a product hidden below the mobile fold, and fixed it before handing the work back. Lovable, the app-building platform, highlighted far less variance run to run alongside better peak performance.

Consistency is the metric that decides whether an automated workflow survives production. A model that scores 95% on a benchmark but swings wildly between runs is worse for automation than one that scores 90% every time. You can build guardrails around a predictable model. You can't build anything durable on a volatile one.

Two quiet updates for orchestration builders

Alongside the model, Anthropic shipped two beta features that got almost no headlines but solve real production problems.

The first is mid-conversation tool changes. Developers can now change which tools Claude has access to during a conversation without invalidating the prompt cache. Complex workflows often need different tools at different stages, and until now, swapping them meant breaking the cache and paying to reprocess context you'd already paid for. This removes a real tax on sophisticated pipeline design.

The second is automatic fallbacks. Requests flagged by safety classifiers can now route to another model automatically instead of being blocked. For production systems, a blocked request is a broken workflow and usually a page to an on-call engineer. A rerouted request is a log entry. That's the difference between a system you babysit and a system you trust.

What this means for your automation strategy

A few practical takeaways if you're building or scaling AI workflows right now.

Audit your workflows by difficulty, not by department. The effort dial only saves money if you know which tasks are actually easy. Most teams have never classified their automated tasks this way. Start there. You'll likely find that a majority of your AI workload is routine work currently running on expensive default settings.

Treat cost predictability as a requirement, not a bonus. When you evaluate models or platforms, ask about variance, not just averages. What does the 95th percentile run cost? What happens when a task fails? The answers matter more than benchmark scores.

Price in the failure rate. A model that costs 20% more per run but fails half as often is usually the cheaper model. Do the math on your own workflows before defaulting to the lowest per-token price.

Watch where the market is going. Anthropic's launch charts emphasize performance at a given cost rather than peak performance alone, and OpenAI has made similar moves around compute controls. When every major lab starts competing on economics, the buying criteria have changed. Capability got commoditized. Efficiency is the new frontier.

The bottom line

Claude Opus 5's significance isn't in what it can do. It's in what it signals: the era of treating AI costs as an unpredictable force of nature is ending. Intelligence is becoming something you provision, tune, and budget like any other infrastructure.

For teams building automated workflows, that's the best news in a long time. The hard part was never getting AI to do impressive things in a demo. It was getting it to do useful things a thousand times a day at a cost you could defend in a budget meeting. We're finally getting the tools for that job.