Anthropic drops Opus 5 at half of Fable 5’s price

Anthropic released Claude Opus 5 on Friday, July 24. The company sold it to developers and enterprise buyers as a model somewhere around Fable 5’s intelligence for half the price.

It follows Opus 4.8 by two months, and gives cost-conscious teams a cheaper default than the flagship they’ve been paying a premium to use.

Opus 5 matches Fable 5 at half the price

Opus 5 has the same operating cost as its predecessor. But it does a lot more, Anthropic says. The model outperforms all competitors on the coding assessment Frontier-Bench v0.1. It is cheaper per task, and outperforms Opus 4.8 by more than a factor of two. There’s a tradeoff. An effort setting puts buyers in control. Buyers can choose between maximum intelligence or dialling down to spend fewer tokens on faster, cheaper answers.

It’s also now the default model on Claude Max, and the most capable option on Claude Pro. Anthropic sells it as a workhorse for everyday use, not as a specialist. The company claims it is more efficient than the alternatives sitting next to it.

On the CursorBench 3.2 coding test, Opus 5 at maximum effort is within half a percentage point of the best score from Fable 5, at half the per-task cost. The smaller of the two is the Opus 5. But it beats Fable 5 outright on several of the benchmarks Anthropic released.

On the novel problem test ARC-AGI 3, Opus 5 was three times better than the second best model. The pass rate on Zapier’s AutomationBench was approximately 1.5 times higher than the nearest competitor, for the same price. It also beat Fable 5’s best score in the OSWorld 2.0 computer use benchmark, for a third of the price.

Anthropic admitted that in cybersecurity work, Opus 5 lags behind Mythos 5.

Fewer restrictions add a new fallback option

Opus 5 falls outside of the 30 day data retention policy that is in place for Fable and Mythos. Anthropic expects its safety classifiers to fire 85% less on Opus 5 than on Fable 5.

But there are still guardrails for security tasks. Opus 5 cannot scan a compiled software binary for vulnerabilities. It can still search source code for them, since source level work is seen as more defensive. Anthropic is rolling out a beta feature called Automatic Fallbacks to reduce the pain when a classifier fires. It then routes a flagged request to a weaker model, ensuring that API users get a working answer instead of an error.

Anthropic gave examples of the model working around obstacles, not stopping at them. For one task on the Frontier Bench, the model was shown a drawing of a machine part and asked to reproduce it as a 3D FreeCAD model. It couldn’t see the photo. Opus 5 wrote its own computer vision code to read the geometry from raw pixels and then reconstructed the part. No rival model succeeded after five tries.

Devin CEO Scott Wu quoted Anthropic’s launch materials in saying the model “approaches Fable-level performance at half the cost” on the firm’s FrontierCode 1.1 test. Zapier CEO Wade Foster said that Opus 5 topped his company’s AutomationBench leaderboard. It also achieved a 100% pass rate on an end-to-end churn prevention workflow that earlier models had failed. In life sciences work, Anthropic reported a 10.2 percentage point increase over Opus 4.8 in inferring molecular structures from spectroscopy data.

Mythos 5, Fable 5, and Sonnet 5 shipped back in June. Haiku is the only model still on the older generation. Cryptopolitan reported the company reached a ~$350 billion valuation in November 2025. The company generates about 80% of its revenue from over 300,000 business and enterprise customers.