Claude Opus 5: Near-Fable Performance at Half the Cost

Anthropic has released its strongest everyday model yet, and crucially, it costs the same as what it replaces.

Claude Opus 5 went live on 24/07/2026. Sitting between Sonnet 5 and Fable 5 in Anthropic's model hierarchy, it is priced at $5 per million input tokens and $25 per million output tokens, the same as Opus 4.8. The performance improvement, however, is substantial. On Frontier-Bench v0.1, which measures software engineering capability, Opus 5 is now the top-ranked model and more than doubles Opus 4.8's score at a lower cost per task.

On CursorBench 3.2, which tests coding across realistic developer tasks, Opus 5 comes within 0.5% of Fable 5's peak score at its highest effort setting while costing half as much per task. That is not a marginal difference; it represents a genuine shift in where the price-to-performance sweet spot sits across Anthropic's model family.

Business automation

For teams using AI to automate multi-step business tasks, the Zapier AutomationBench results are the most telling. Opus 5's pass rate is approximately 1.5 times that of the next-best model at the same cost per task. Zapier's CEO confirmed in early-access testing that Opus 5 completed a full churn-prevention sequence end to end, flagging at-risk accounts, alerting the right account owner, and producing a retention summary, while previous models failed the same task. Even at its lowest effort setting, Opus 5 passes more tasks than any competing model at higher settings.

This matters for UK businesses running agentic workflows because failed tasks generate retried requests, which means wasted tokens and longer completion times. A model that gets it right first time is meaningfully cheaper to operate than a cheaper model that needs multiple runs.

The effort setting

One feature worth understanding is Opus 5's effort setting. Rather than simply picking a model and running it, developers can tune how much reasoning Opus 5 applies to a task. At lower effort it is faster and more economical; at higher effort it is thorough but slower and more token-intensive. Anthropic's benchmarks show Opus 5 still outperforms rivals even at minimum effort, which gives teams useful flexibility: lighter effort for routine tasks, heavier effort where depth matters.

A Fast mode is also available at approximately 2.5 times the default speed, priced at double the base rate, so $10 per million input tokens and $50 per million output tokens. That suits latency-sensitive workflows where response time carries a real business cost.

Safety and alignment

Anthropic reports Opus 5 is their most aligned model to date, scoring 2.3 on an internal misaligned behaviour audit, below Opus 4.8, Sonnet 5, and Fable 5. It also shows the lowest rates of deceptive behaviour and is the least susceptible to being manipulated into misuse.

On cybersecurity, the model applies lighter restrictions than Fable 5, allowing source-code vulnerability analysis but still blocking penetration testing, binary-based scanning, and exploit generation without enrolment in Anthropic's Cyber Verification Programme. Requests that trigger classifiers automatically fall back to Opus 4.8.

What this means in practice

The migration path is straightforward for teams already using Opus 4.8 via the API: switch the model identifier to claude-opus-5 and the pricing is unchanged, so there is no cost risk in running your existing evaluation suite first.

If you are on Claude Pro, Opus 5 is now the strongest model on offer. If you are on Claude Max, it is the new default.

Two areas stand out for UK businesses specifically. The first is agentic coding: Opus 5 verifies its own work, holds context across long sessions, and iterates more carefully before delivering a result, which reduces the back-and-forth that typical coding agents require. The second is document and financial analysis: early enterprise testing by Box found an 11% improvement in data analysis tasks and a 17% improvement in due diligence workflows compared to Opus 4.8.

For compliance-conscious teams, it is also worth noting that Opus 5 carries no data retention requirements for general access, consistent with prior Opus models.

At Adevious AI

We have been running Opus 5 since launch and the gains on sustained, context-heavy tasks are clear. For the CoWork agent workflows and Claude Code integrations we build for clients, Opus 5 is now the practical default for anything that benefits from deep reasoning across multiple steps. The effort setting also adds useful control: we can tune token spend to fit the complexity of each task rather than applying maximum reasoning to every call.

If you want to understand where Opus 5 fits into your current AI setup, or whether now is a good time to revisit the models powering your workflows, Adevious AI is happy to help. Get in touch at adevious.co.uk.