Claude Opus 5 launches as Anthropic’s lower-cost model for Claude users
Anthropic says Claude Opus 5 approaches Mythos 5 quality in many areas while cutting cost and tightening safety controls for developers.
By Wei-Lin Zhao · AI Correspondent
· 3 min read
Anthropic has launched Claude Opus 5 for its Claude chatbot and developer platform, positioning the new large language model as a cheaper alternative that approaches the output quality of its unreleased flagship Mythos 5 in many areas. The company did not disclose model training costs, revenue impact or customer adoption targets.
According to Anthropic, Opus 5 is now available through Claude’s top-end Max tier as the default model and is also offered in the Pro plan. Developers can access it through Anthropic’s API.
How does Claude Opus 5 compare with Fable 5?
Anthropic has kept Mythos 5 out of public release because, it says, the model could be used to find software flaws and build exploitation workflows. Instead, the company offers Fable 5, a smaller version that shares many Mythos 5 capabilities while presenting lower cybersecurity risk.
Anthropic said it tested Opus 5 against Fable 5 on 13 benchmarks. Opus 5 beat Fable 5 on eight of those tests while costing about 50% less, according to the company. The performance gap was largest on Frontier-Bench and AutomationBench, two tests that Anthropic cited in its announcement.
- On Frontier-Bench, a 74-task benchmark covering areas including physics, chemistry and cryptography, Anthropic said Opus 5 scored 9.7% higher than Fable 5.
- On AutomationBench, which measures knowledge-work tasks, Anthropic said Opus 5 scored 8.5% higher.
- Anthropic said Opus 5 outperformed Opus 4.8 on all of its internal life-sciences benchmarks, including automated protein research tasks.
Those results are company-reported, and Anthropic did not provide enough public detail to compare Opus 5 independently with rival frontier models. The useful signal for enterprise buyers is narrower: Anthropic is trying to move more capability into a model it is willing to sell broadly, while keeping its highest-risk model unavailable.
What is new in Opus 5?
Anthropic said one design change in Opus 5 is its ability to check its own work. In a Frontier-Bench task requiring a three-dimensional machine-part schematic, the model could not directly inspect the drawing it produced. Anthropic said Opus 5 responded by building a computer-vision workflow to extract geometry from the image pixels and assess the result.
The company is also leaning on that verification behavior for scientific workloads. Anthropic said Opus 5 performed especially well in automated protein research, a field that drew broader attention after the 2024 Nobel Prize in chemistry recognized work in that area.
Safety is the other major claim. Anthropic described Opus 5 as its safest model so far across several measures, saying it is less likely than earlier models to engage in deceptive behavior or make errors that are hard to undo. The company also said Opus 5 received a zero score on an internal test measuring a model’s ability to exploit vulnerabilities.
Anthropic is loosening some restrictions compared with Fable 5. Customers can use Opus 5 to look for security problems in code, but the model is designed to block vulnerability-scanning techniques commonly associated with attackers, according to the company. Anthropic estimates those guardrails will trigger 85% less often than Fable 5’s controls.
The API update shipping with Opus 5 adds two developer-facing controls. One lets developers stop a model during a workflow and change the tools available to it. Another can send prompts blocked by Anthropic’s safety systems to a different model. Anthropic did not say how often that fallback will be used or which workloads it expects to shift first.
This story draws on original reporting from SiliconANGLE.
