Anthropic ships Claude Opus 5: near-Fable-5 intelligence at half the price

On July 24, 2026, Anthropic announced Claude Opus 5, the latest model in its Opus family, built to deliver nearly the full intelligence of the flagship Claude Fable 5 at exactly the same price as its predecessor, Opus 4.8. It is the company's fourth new model in under two months, following Mythos 5, Fable 5, and Sonnet 5 in June.

Opus 5 is not another minor version-number bump. Anthropic is positioning it directly as "near-Fable-5 intelligence at half the cost" - at an unchanged price versus its predecessor.

The facts: pricing, availability, positioning

The biggest change here isn't the price itself, it's what you get for it. Anthropic claims Opus 5 approaches Fable 5's level - the model that has served as the company's flagship for raw intelligence - while staying in the cheaper, faster Opus tier. For API users, that means a model that genuinely competes with Anthropic's priciest offering, without raising the bill.

Benchmarks: what actually improved

On ARC-AGI 3, a test measuring the ability to solve genuinely novel, non-templated problems, Anthropic reports a score three times higher than the next-best model. That's a large jump for a single update, and this particular benchmark is specifically designed to resist being "gamed" by training on similar tasks - it measures reasoning flexibility rather than pattern memorization.

On GDPval-AA v2, a benchmark combining coding with knowledge work, and on Frontier-Bench v0.1, Opus 5 sets new state-of-the-art results among Anthropic's models, doubling Opus 4.8's score on the latter. On OSWorld 2.0, a computer-use benchmark (navigating interfaces, clicking, filling out forms), the model surpasses Fable 5's best result at roughly a third of Fable 5's cost. On CursorBench 3.2, a coding-assistance test, Opus 5 lands within half a percentage point of Fable 5's score while costing half as much per task.

The new effort parameter

Opus 5 introduces an effort setting with four levels: low, medium, high, and extra-high. Higher effort means the model spends more tokens reasoning before answering, which raises both quality and cost. This lets developers deliberately choose between a fast, cheap answer to a simple question and deeper, costlier reasoning where it's actually warranted, instead of paying full rate on every request regardless of how complex it actually is.

Safety and alignment

Anthropic describes Opus 5 as its most aligned model to date according to its internal behavioral audit, with a significantly lower misalignment score than prior versions. The company also reports roughly 85% fewer cybersecurity-related safety classifier interventions compared to Fable 5, suggesting fewer false positives for legitimate security-related work.

At the same time, Opus 5 remains behind Mythos 5 in resistance to attempts to use it for offensive cybersecurity work and high-risk biological research - so Anthropic maintains a clear hierarchy of models by restrictiveness in the most sensitive domains, rather than applying one universal threshold across the entire Opus family.

Why this is the fourth model in two months

Opus 5 is Anthropic's fourth model release in under two months, counting from Mythos 5, through Fable 5, to Sonnet 5 in June. That pace stands in sharp contrast to the company's earlier release cadence, where gaps between major models were measured in quarters, not weeks. Part of the explanation for that speed sits in the business context: Anthropic is preparing for a planned IPO later this year, and a run of fast, concrete model releases is a natural way to demonstrate continuous technical progress to the market right before a public debut.

What this means for users and developers

For someone on Claude Pro or Claude Max, the change is essentially immediate and free - Opus 5 replaces the previous model as the default or strongest available option, with no extra configuration needed. For developers on the API, the shift is just as simple: the model identifier changes to "claude-opus-5", pricing stays the same, and the new effort parameter gives finer control over the cost of individual requests than was previously available in this form.

SPAWNSY verdict