Anthropic announced Claude Opus 5 on July 24, 2026, describing it as a model for long-running agents, software development, knowledge work and professional workflows. The company says Opus 5 outperforms its predecessor, Opus 4.8, on coding, business-task, computer-use and life-sciences evaluations while retaining the same base token prices.

The results come from Anthropic’s own announcement and pre-deployment evaluations rather than an external audit or peer-reviewed study. The company also says Opus 5 remains behind Mythos 5 on some autonomous biology and offensive-cybersecurity capabilities.

Contents

What changed

Anthropic positions Opus 5 as a successor to Opus 4.8 with higher performance at the same standard token prices:

  • $5 per million input tokens
  • $25 per million output tokens

Input tokens are the text or other information supplied to the model, while output tokens are the content it generates. These prices do not by themselves determine the cost of a completed task: long-running agents may consume many model calls, use external tools and generate large amounts of output.

The model is available through the Claude API under the identifier claude-opus-5. Anthropic also offers a Fast mode that runs at about 2.5 times the default speed. Fast mode costs twice the base price on the Claude Platform and through usage credits in Claude Code.

The launch adds two beta features. Developers can change tools during a conversation without invalidating the prompt cache, and Anthropic can automatically route some API requests flagged by its classifiers to another model. In Claude.ai, Claude Code and Claude Cowork, certain flagged requests may fall back to Opus 4.8.

Anthropic also says Opus 5 has no data-retention requirements for general access, consistent with its prior Opus models.

What Anthropic reports from its evaluations

Anthropic says Opus 5 is state of the art on its coding and knowledge-work evaluations, including Frontier-Bench and GDPval-AA, while remaining behind Mythos 5 on cybersecurity tasks. The announcement provides headline comparisons for several evaluations but does not provide the numerical scores, full task sets or complete methodologies for most of them.

On the company’s internal Frontier-Bench v0.1 software-engineering evaluation, Anthropic says Opus 5 more than doubled Opus 4.8’s performance at a lower cost per task. The reported run used the mini-SWE-agent harness and a Google Kubernetes Engine backend, with the mean reward calculated over five attempts per task.

Anthropic also reports that Opus 5:

  • came within 0.5% of Fable 5’s peak score on CursorBench 3.2 at maximum effort, while costing half as much per task;
  • scored three times as high as the next-best model on ARC-AGI 3;
  • achieved about 1.5 times the next-best model’s pass rate on Zapier AutomationBench at the same cost per task;
  • outperformed every other model at any given cost on OSWorld 2.0; and
  • exceeded Fable 5’s best OSWorld 2.0 result at slightly more than one-third of the cost.

These comparisons depend on the evaluation setup. Anthropic says some results use adjustable effort settings. Higher effort generally allows a model to spend more computation on a task, potentially improving results while increasing latency or token use. A score at maximum effort therefore should not automatically be treated as a fixed property of every deployment.

The company also reports stronger performance on its life-sciences evaluations. Opus 5 outperformed Opus 4.8 across all of the company’s reported life-sciences tests, including a 10.2 percentage-point improvement on an organic-chemistry benchmark and a 7.7 percentage-point improvement on a protein-related benchmark.

Those tests cover tasks such as inferring molecular structures from spectroscopy data and reasoning about protein-sequence variation. They are model evaluations, not evidence that Opus 5 can independently conduct validated scientific research or replace laboratory confirmation.

Anthropic describes early-use examples in which Opus 5 reconstructed a machine part as a 3D FreeCAD model, fixed an underlying package-manager bug, built a market-data feed and created a test harness. The examples illustrate the type of multi-step work Anthropic is targeting, but customer and individual-task examples cannot establish general reliability.

Why agentic performance is different from a single answer

An AI agent is a model-based system that can plan across several steps, use external tools, inspect the results and revise its actions. It may need to maintain context, test its own output, recover from an error and decide what to do next.

That makes agent performance different from producing a correct response to a single prompt. A coding agent, for example, may need to:

  1. understand an existing codebase;
  2. identify the relevant files;
  3. make a change;
  4. run tests;
  5. interpret failures;
  6. revise the implementation; and
  7. confirm that the final result works.

Anthropic’s Opus 5 examples emphasise this iterative pattern, including creating test harnesses and checking work. If the reported benchmark results transfer to production systems, stronger performance at the same base list price could make longer coding, research and business workflows more practical.

However, the overall cost depends on more than the model’s token rate. Tool calls, repeated attempts, context size, effort setting, latency and the cost of handling failures can all affect the price of a completed task. The announcement does not provide equivalent configuration, latency or token-budget details for most of its cost-per-task comparisons.

Cybersecurity boundaries and remaining limitations

Anthropic separates vulnerability discovery from exploit development. Finding a potential weakness in source code and reliably producing a working exploit are distinct capabilities, so performance in one area does not establish equivalent performance in the other.

The company says Opus 5 is close to Mythos 5 at identifying vulnerabilities but substantially behind it at developing exploits. Anthropic also says it did not intentionally train Opus 5 on cyber tasks.

In the standard product experience, Anthropic’s cybersecurity safeguards permit vulnerability finding in source code but block or restrict:

  • binary-based vulnerability scanning;
  • penetration testing; and
  • exploit generation.

Anthropic expects the Opus 5 cyber classifiers to intervene about 85% less often than the classifiers used for Fable 5. This is a company estimate, not an independently verified measurement in the supplied evidence. Users working on cyber tasks may still encounter classifier restrictions, fallback behaviour or eligibility requirements for Anthropic’s Cyber Verification Program.

Anthropic reports an overall misaligned-behaviour score of 2.3 in its automated behavioural audit, which it says was the lowest among the recent models it compared. The announcement does not define the audit’s full composition or statistical methodology, so the score’s meaning and comparability are difficult to assess without the underlying evaluation details.

The company says Opus 5 also remains behind Mythos 5 in biology research and offensive cybersecurity, and acknowledges important limitations on long-running autonomous biology tasks.

Pricing and access

Opus 5 is available immediately through the Claude API as claude-opus-5, at the same base token prices Anthropic lists for Opus 4.8: $5 per million input tokens and $25 per million output tokens.

It is also positioned for use through Claude.ai, Claude Code and Claude Cowork. Anthropic says the model becomes the default on Claude Max and the strongest model on Claude Pro.

The launch does not establish India-specific pricing, taxes, payment methods, availability or regulatory status. Those details would need to be verified separately.

What the evidence does not establish

The supplied evidence is a company product announcement describing Anthropic’s own benchmarking, safety evaluations and early-access customer testing. It does not provide an independent replication of the results.

Several limitations matter when interpreting the comparisons:

  • numerical scores, confidence intervals and complete task sets are missing for most evaluations;
  • the exact model configurations, token budgets, latency measurements and tool environments are not specified for most cost comparisons;
  • Frontier-Bench reports a mean reward over five attempts per task, while comparable repeated-trial information is not provided for most other benchmarks;
  • it remains unclear how the results transfer to different agent frameworks and production workloads;
  • an automated behavioural score does not by itself establish safety in every deployment;
  • customer examples and testimonials cannot establish general reliability; and
  • the announcement does not show that Opus 5 is suitable for unsupervised decisions in software, finance, law, scientific research or security.

The next important evidence will be independent replication of the benchmark results, including cost, latency and failure rates, along with more detail in Anthropic’s System Card. Real-world testing will also need to examine whether long-running agents can recover from tool errors, avoid harmful side effects and preserve relevant context over extended workflows.

Sources