A new flagship for everyday agent work

Anthropic has released Claude Opus 5, the first Opus-branded model in its fifth generation. The company presents it as a model for sustained professional use: software engineering, document-heavy analysis, research support and multi-step computer work. It succeeds Claude Opus 4.8 at the same published API price, rather than introducing a new premium tier.

The central commercial argument is efficiency. Anthropic says Opus 5 reaches close to Claude Fable 5’s capability in several coding evaluations while costing about half as much per task. Fable 5 remains the company’s more capable generally available model for the most ambitious, long-duration autonomous projects, but it carries substantially tighter safety controls. Opus 5 therefore occupies an important middle position: intended to make stronger agentic behaviour practical for routine development and enterprise workflows.

That positioning is more nuanced than a simple claim that Opus 5 is universally better than Fable 5. Results vary by evaluation, effort level and task type. Anthropic reports that Opus 5 can exceed Fable 5 on some benchmarks when measured against cost, while Fable 5 retains an advantage on the longest and most autonomous workloads.

Performance claims centre on coding and knowledge work

Anthropic’s published results emphasise software engineering. On Frontier-Bench, an evaluation of complex engineering tasks, the company says Opus 5 more than doubles the performance of Opus 4.8 while reducing cost per completed task. On CursorBench, it says the model comes within 0.5 percentage points of Fable 5’s peak score at roughly half the cost.

The release also highlights results in business automation, computer use and abstract problem solving. Anthropic says Opus 5 is its strongest and most cost-efficient model across a range of internal and external-style evaluations, including tasks involving browser interaction, research and office workflows. It further reports improvements over Opus 4.8 in life-sciences assessments, including molecular-structure inference and protein-related analysis.

These results indicate an effort to shift the comparison away from raw model hierarchy and towards output per unit of compute. For organisations deploying agents at scale, a modest reduction in capability may be acceptable if it is accompanied by lower token consumption, faster responses or more reliable task completion. That calculation is especially relevant for workflows in which a model must repeatedly inspect files, use tools, test code and revise its work.

However, benchmark leadership does not automatically establish equivalent performance in production. Anthropic’s launch materials include company-run tests and reports from early-access customers, which can be useful evidence of adoption but are not independent audits. Enterprises will still need to test the model against their own codebases, data quality requirements, tool permissions and failure tolerances before treating stated gains as operationally proven.

A model designed to check and continue

A prominent theme in Anthropic’s description is improved persistence. The company says Opus 5 is better at verifying intermediate work, revising plans and recovering from errors rather than stopping at the first plausible answer. Such behaviour is central to the current market for AI agents, where value is increasingly tied to carrying a task through several steps with limited human intervention.

The distinction matters in programming. A useful coding agent must not only generate a patch; it must understand a repository, inspect the result, run tests, identify regressions and adapt. Anthropic describes examples in which Opus 5 built supporting tooling to complete an assignment, including a computer-vision pipeline used to reconstruct a component from image data. These examples are demonstrations rather than a guarantee of general performance, but they illustrate the type of autonomy the company is selling.

For users, the practical question is whether the greater thoroughness produces better outcomes without excessive latency or token use. Anthropic offers adjustable effort settings, allowing developers to trade off speed and expense against deeper reasoning. It also offers a faster operating mode, which it says runs about 2.5 times faster at double the base price. The availability of these options reinforces that model selection is now partly a workload-management decision.

Safety separates Opus 5 from Fable 5

Opus 5 arrives after the debut of Fable 5 and Mythos 5 in June 2026 put greater attention on how Anthropic handles advanced capabilities in cybersecurity and biology. Anthropic says Opus 5 does not extend the frontier in the riskiest dual-use areas and remains behind Mythos 5 in offensive cyber and biology research evaluations.

The company says it deliberately avoided training Opus 5 specifically on cyber tasks. Nonetheless, wider improvements in general reasoning have made the model more effective at identifying vulnerabilities. Anthropic reports that Opus 5 is closer to Mythos 5 at finding software flaws than at turning those flaws into working exploits. Its safeguards permit certain defensive activities, such as source-code vulnerability review, while blocking binary-based scanning, penetration testing and exploit development.

Anthropic expects these safeguards to intervene far less often than Fable 5’s controls. When a request is flagged in Claude’s consumer and coding products, it can be routed by default to Opus 4.8 instead. This fallback approach aims to preserve access to ordinary work while restricting selected high-risk actions, but it can also complicate reproducibility: users may need to know which model actually handled a sensitive request.

Price and availability strengthen the competitive case

Claude Opus 5 is available through Anthropic’s platforms and the Claude API at $5 per million input tokens and $25 per million output tokens, matching the published pricing for Opus 4.8. It is also available through Amazon Bedrock. Anthropic made it the default model for Claude Max subscribers and describes it as the most capable option available to Claude Pro users.

Keeping the prior Opus price is strategically significant. It makes the release less about adding a premium flagship than about lowering the cost of advanced agent use. Reuters reported that Anthropic characterises Opus 5 as nearing Fable 5’s capability at half the price, while recommending Fable 5 for projects that require days-long autonomy.

The result is a more segmented Claude portfolio. Fable 5 remains aimed at the furthest-reaching agent tasks under more restrictive controls; Opus 5 becomes the broad workhorse for demanding coding and analytical use; and earlier models remain relevant for compatibility, lower-cost processing or fallback handling. The early evidence supports Anthropic’s claim of a significant improvement over Opus 4.8, but the more consequential test will be whether Opus 5 delivers dependable, economical results in real enterprise systems rather than only in controlled evaluations.

Sources