On July 24, 2026 (US time), Anthropic released Claude Opus 5, a new model for its Claude AI assistant[1]. The company describes it as a model that comes close to the intelligence of its top-tier Claude Fable 5 at half the cost, and says it reached a new state of the art on coding and knowledge-work evaluations. Opus 5 becomes the new default model on the consumer Claude Max plan and is available across all platforms from launch day[1].
Near-Fable 5 Intelligence at a Lower Cost
Anthropic positions Claude Opus 5 as a high-performance model built for everyday use. According to the company, it delivers a major jump in performance at the same price as its predecessor, Opus 4.8, and an effort setting lets users choose between prioritizing intelligence or conserving tokens for faster, cheaper results[1].
On coding and knowledge-work evaluations such as Frontier-Bench and GDPval-AA, Opus 5 sets a new state of the art. The company is explicit, however, that on cybersecurity tasks it still trails its own Mythos 5 model[1]. Notably, Anthropic separates out where the model is and is not strong rather than overselling it.
Topping the Major Benchmarks
Anthropic also published concrete numbers[1]. On Frontier-Bench v0.1, which measures software engineering, Opus 5 more than doubles Opus 4.8's performance while actually lowering the cost per task. On the CursorBench 3.2 code-editing benchmark, at maximum effort it lands within 0.5% of Fable 5's peak score at half the cost per task.
On ARC-AGI 3, which tests solving novel problems, Opus 5 scored three times as high as the next-best model. On Zapier AutomationBench, which checks whether a model can complete business tasks end to end, it reached about 1.5 times the pass rate of the next-best model at the same cost, and even at its lowest effort setting it cleared more tasks than any other model. On the OSWorld 2.0 computer-use benchmark, it outperformed every model at any given cost and surpassed Fable 5's best result at roughly one-third of the cost[1].
Verifying Its Own Work as It Goes
Anthropic says Opus 5 excels at checking its own work and iterating until it succeeds[1]. In one evaluation, the model was asked to rebuild a machine part as a 3D model in code from a drawing, but was given no direct way to view the drawing. Opus 5 responded by writing its own computer-vision pipeline, reading the geometry from the raw pixels and reconstructing the part. In another case, it traced the root cause of a bug in a widely used package manager that the community's own patch had missed[1].
There is progress in scientific research as well. Opus 5 beat Opus 4.8 across every life-sciences evaluation, scoring 10.2 points higher on an internal organic-chemistry task that infers molecular structures from spectroscopy data, and 7.7 points higher on protein-related tasks[1].
"Most Aligned" on Safety, With Cyber Capability Held Back on Purpose
Anthropic also released safety results. In pre-deployment testing, it calls Opus 5 its most aligned model to date, saying it adheres to its guiding Constitution better than Opus 4.8, Sonnet 5, or Fable 5, and is the most resistant to deceptive behavior and to being tricked into misuse[1].
At the same time, potentially dangerous capabilities are deliberately restrained. The model trails Mythos 5 in biology research and offensive cybersecurity, and Anthropic says it intentionally avoided training on cyber tasks. Opus 5 comes close to Mythos 5 at finding vulnerabilities but remains far behind at turning them into real threats through exploitation[1]. Its safety classifiers are expected to intervene about 85% less often than those on Fable 5, allowing source-code vulnerability research while blocking penetration testing and the generation of exploit code[1].
Availability and Pricing
Claude Opus 5 is available on all platforms from launch day. Pricing is 5 USD (about 820 yen) per million input tokens and 25 USD (about 4,100 yen) per million output tokens, unchanged from Opus 4.8[1]. Developers can use it as claude-opus-5 on the Claude API, and a Fast mode that runs about 2.5 times faster is offered at twice the base price[1].
※1 USD = 164 JPY (as of July 24, 2026)
Two beta features arrive alongside the model. On the Claude Platform, developers can now change which tools are available mid-conversation without invalidating the prompt cache, and on the API an automatic fallback can reroute requests flagged by the safety classifiers to another model[1].
Summary
Claude Opus 5 is a new model built around a simple pitch: intelligence approaching the top-tier Fable 5 at half the cost. It posts state-of-the-art results across many benchmarks while deliberately keeping dangerous capabilities such as cyberattacks and advanced biology research below Mythos 5, and it emphasizes strong alignment. Holding the price steady from Opus 4.8 while lifting performance, it looks like a move to sharpen Claude's everyday workhorse model.
