Anthropic has released Claude Opus 5, a new premium AI model aimed at demanding coding, computer-use and knowledge-work tasks. Launched on 24 July 2026, it promises performance close to the company’s frontier Claude Fable 5 model at half the price, while retaining the previous Opus model’s API pricing.
That combination makes this more than another benchmark-driven model update. For developers and businesses, Claude Opus 5 could offer a practical middle ground: stronger autonomous task completion and self-checking without paying for Anthropic’s most expensive tier.
Claude Opus 5 at a glance
- Availability: Claude.ai, Claude Code, Claude Cowork and the Claude API
- API model name:
claude-opus-5 - Price: US$5 per million input tokens and US$25 per million output tokens
- Fast mode: About 2.5 times the default speed at twice the base API price
- Best suited to: Complex coding, research, business automation, computer use and multi-step analysis
Background: where Opus 5 fits in Anthropic’s model range
Anthropic positions Opus as a high-capability model for difficult everyday work. Claude Opus 5 sits below the more powerful Fable 5 in the company’s range, but is intended to deliver much of that model’s intelligence more efficiently.
It replaces Opus 4.8 at the same token price. Anthropic says Opus 5 is now the default model for Claude Max subscribers and the strongest option available on Claude Pro. For API customers, the unchanged price means teams can test it against existing Opus workflows without first accepting a higher per-token rate.
What changed with Claude Opus 5?
Stronger agentic coding and computer use
The central improvement is not simply generating better answers. Anthropic says Opus 5 is more capable of checking its own work, identifying mistakes and iterating until a task succeeds. That behaviour is important for AI agents, which may need to inspect files, operate tools, test code and recover from failed steps with less human intervention.
In Anthropic’s published evaluations, Opus 5 led Frontier-Bench v0.1 for software-engineering tasks. At maximum effort, it reportedly came within 0.5 percentage points of Fable 5’s peak CursorBench 3.2 score while costing half as much per task. It also recorded strong results on OSWorld 2.0, a benchmark designed to measure computer-use capability.
Benchmarks are not the same as production performance, and several results come from Anthropic’s own testing. Even so, the emphasis on root-cause analysis, verification and long-running tasks points to where the model may be most useful: substantial work that extends beyond a single prompt and response.
Adjustable effort for cost and speed control
Claude Opus 5 supports effort settings that let customers trade additional reasoning and token use for higher performance. A developer could reserve maximum effort for complex migrations or difficult debugging, while selecting a lower setting for routine analysis.
This matters because the cheapest model per token is not always the cheapest model per completed task. A more capable system may finish work in fewer attempts, with fewer tool calls and less human correction. Teams should therefore compare total workflow cost and success rate rather than API pricing alone.
New API workflow features
Anthropic also introduced two beta capabilities alongside the model. Mid-conversation tool changes allow developers to alter the tools available to Claude without invalidating the prompt cache. This can help applications expose only the tools needed at each stage of a workflow.
Automatic fallbacks can route a request to another model if Opus 5’s safety classifier blocks it. Instead of returning only an error, an application may still receive a usable response from an appropriate alternative model. Developers need to enable and test this behaviour carefully, because a fallback model may differ in capability, cost and output style.
Why Claude Opus 5 matters
The launch reflects a broader shift in the AI market from maximising benchmark scores at any cost to improving useful intelligence per dollar. Businesses increasingly need models that can complete repeatable work reliably, integrate with tools and fit within predictable budgets.
For software teams, the practical opportunity is deeper assistance with debugging, code review, refactoring and feature implementation. For analysts and operations teams, Opus 5 may be useful for multi-step research, spreadsheet work and automations that span several systems. Anthropic also reports improvements across internal life-sciences evaluations, although scientific conclusions still require expert review and independent validation.
Practical impact for developers and businesses
Existing Opus 4.8 users have the clearest reason to run an evaluation. Because base pricing is unchanged, they can compare task completion, latency, token use and human-review time using a representative set of real jobs.
A sensible pilot should include:
- Tasks with known correct outcomes and clear acceptance tests
- Measurement of end-to-end cost, not only token consumption
- Separate testing at low, high and maximum effort settings
- Human approval for consequential code, financial or legal outputs
- Monitoring for tool errors, unexpected fallbacks and inconsistent results
Teams should avoid replacing a production model solely because a launch chart looks favourable. Internal data, tool permissions and failure modes often matter more than a small benchmark difference.
Risks, limitations and safety concerns
Claude Opus 5 remains an AI model and can produce incorrect claims, flawed code or overconfident recommendations. Better self-verification may reduce some mistakes, but it does not guarantee correctness. Autonomous workflows need permission limits, audit logs, testing and a clear path for human intervention.
Anthropic says Opus 5 does not advance the frontier in high-risk dual-use capabilities and remains behind Mythos 5 in offensive cybersecurity and some biological research. The company allows source-code vulnerability discovery but applies stronger restrictions to exploit generation, penetration testing and binary-based vulnerability scanning.
Those safeguards are expected to intervene substantially less often than Fable 5’s classifiers. That may improve usability for legitimate work, but organisations should still confirm that their security, privacy and regulatory requirements are met before sending sensitive information to any hosted model.
What to watch next
The most important evidence will come from independent tests and production deployments. Watch for comparisons of Opus 5 with Fable 5 and competing frontier models on real coding repositories, computer-use reliability, latency and total cost per successful task.
Developers should also monitor the beta tool-change and automatic-fallback features. If they prove reliable, they could make agentic applications more efficient and less brittle. Anthropic may refine pricing, limits and model routing as usage grows.
Conclusion
Claude Opus 5 appears designed for organisations that want frontier-adjacent capability without frontier-tier cost. Its unchanged Opus pricing, adjustable effort and focus on verification make it a compelling upgrade candidate for demanding coding and knowledge workflows.
The right response is not an immediate system-wide migration, but a disciplined pilot. If Opus 5 completes more real tasks with fewer retries and less oversight, its value could extend well beyond the benchmark headline.