Anthropic has released Claude Opus 5, a new general-purpose AI model positioned as a more economical alternative to its frontier Claude Fable 5 system. Available across Anthropic’s platforms, Opus 5 costs $5 per million input tokens and $25 per million output tokens, matching the price of its Opus 4.8 predecessor while delivering higher performance on software engineering, knowledge work and scientific tasks.
Opus 5 is now the default model for Claude Max subscribers and the most capable option offered through Claude Pro. Developers can access it through the Claude API under the model name claude-opus-5. Anthropic is also offering a Fast mode that runs about 2.5 times as quickly as the standard configuration, at twice the base API price or through usage credits in Claude Code.
The launch emphasizes the trade-off between model capability and operating cost. Anthropic says Opus 5 approaches Fable 5’s performance at half the price, with results varying by task and the selected effort setting. Those controls let customers allocate more computation to difficult prompts or reduce token use when speed and cost are more important.
In Anthropic’s Frontier-Bench v0.1 software engineering evaluation, Opus 5 recorded the company’s strongest result and more than doubled Opus 4.8’s performance while costing less per task. At its maximum effort setting on CursorBench 3.2, it came within 0.5% of Fable 5’s peak result at half the cost per task. Benchmark results remain dependent on the test harness, configuration and scoring method, and Anthropic disclosed that Opus 4.8 handled requests refused by safety classifiers in some internal runs.
The model also improved across Anthropic’s life-sciences tests, which cover areas such as structural biology, organic chemistry and bioinformatics. It scored 10.2 percentage points above Opus 4.8 on an internal task involving the inference of molecular structures from spectroscopy data, and 7.7 points higher on a test of how protein-sequence changes affect function. Anthropic nevertheless describes Opus 5 as limited on long-running autonomous research, an area where Mythos 5 remains more capable.
Cost has become a central consideration for businesses deploying generative AI, particularly when applications process large volumes of text or repeatedly call a model as part of an automated workflow. Anthropic product executive Dianne Penn told CNBC that customers are seeking value but still require comparable quality.
“Enterprises … are looking for value,” Penn said.
Opus 5 does not replace Mythos 5 as Anthropic’s leading model for sensitive dual-use work. The new model can identify software vulnerabilities at a rate closer to Mythos 5, but remains substantially weaker at turning those findings into working exploits. Anthropic says it deliberately avoided training Opus 5 on cyber tasks, although improvements in its general reasoning also raised its cybersecurity performance.
Safeguards permit source-code vulnerability analysis while restricting binary vulnerability scanning, penetration testing and exploit generation. Anthropic expects the cyber classifiers to intervene about 85% less frequently than those used with Fable 5. Flagged requests in Claude’s consumer and coding products can fall back to Opus 4.8, while approved participants in Anthropic’s Cyber Verification Program can use a less restricted version.
Anthropic also reports that Opus 5 followed its behavioral rules more reliably than Opus 4.8, Sonnet 5 and Fable 5 during pre-deployment testing, while showing lower rates of deceptive behavior and resistance to misuse. The company says general access carries no data-retention requirement, consistent with previous Opus releases.
Sources: anthropic.com, cnbc.com