Anthropic has launched Claude Opus 5, a new artificial intelligence model designed for software development, professional knowledge work and complex tasks that require an AI agent to operate over multiple steps.
The model became available on July 24, 2026, across Anthropic’s platforms. It is now the default model for Claude Max subscribers and the most capable model available through the Claude Pro plan.
Anthropic describes Opus 5 as a more efficient alternative to its most advanced models, claiming that it approaches the performance of Claude Fable 5 at approximately half the price. The company is positioning it as a model that can be used regularly rather than reserved only for the most expensive or difficult tasks.
Claude Opus 5 costs $5 per million input tokens and $25 per million output tokens through the Claude API, the same standard price previously charged for Opus 4.8.
What Is Claude Opus 5?
Claude Opus 5 is the latest model in Anthropic’s Opus family, which is generally aimed at developers, professionals and businesses that need stronger reasoning than smaller or faster models can provide.
The model is designed to work on tasks such as:
- Building and modifying software projects.
- Investigating difficult bugs.
- Reviewing large codebases.
- Analyzing documents and financial information.
- Conducting scientific and business research.
- Operating tools and computer interfaces.
- Completing long, multi-step workflows.
- Checking and correcting its own work.
Instead of focusing only on producing a good first answer, Anthropic says Opus 5 is better at verifying its progress, identifying incomplete work and continuing to iterate until the requested result is achieved.
Stronger Performance for Coding Agents
Software development is one of the main areas emphasized in the Opus 5 announcement.
According to Anthropic’s evaluations, the model performed significantly better than Opus 4.8 on software engineering tasks. The company says Opus 5 more than doubled its predecessor’s result on one version of Frontier-Bench while completing tasks at a lower average cost.
On CursorBench, Anthropic says the model came within 0.5% of Fable 5’s peak score when using its maximum effort setting, while costing approximately half as much per task.
These results come from Anthropic and its partners, so they should not be treated as a complete substitute for independent testing. However, they suggest that Opus 5 is intended to compete directly for complex programming workflows rather than only generate isolated code snippets.
Better at Finding the Root Cause of Bugs
One of the most useful improvements described by Anthropic is stronger root-cause analysis.
AI coding tools sometimes fix the visible symptom of an error without identifying the underlying problem. That can produce a solution that appears to work temporarily but fails again in a slightly different situation.
During one Anthropic evaluation, Opus 5 was given a real bug from a popular open-source package manager. The model reportedly identified the underlying cause and fixed an additional edge case that had been missed by an existing community patch.
For developers, this type of behavior could be especially valuable when investigating:
- Intermittent application failures.
- Unexpected database behavior.
- Authentication and session problems.
- Payment integration errors.
- Build and dependency failures.
- Bugs involving several connected files.
Human review remains necessary, but a model that investigates the real cause instead of applying the first available correction can reduce repeated work.
Designed for Longer Agentic Workflows
Claude Opus 5 is also designed to maintain its focus across longer tasks.
An agentic workflow may require the model to inspect files, create a plan, execute commands, modify code, run tests and revise the implementation several times before finishing.
Anthropic says Opus 5 is more consistent during these longer workflows and more likely to verify the final result before reporting that a task is complete.
In one example shared by the company, the model built a market data feed for a new exchange. When it discovered that no live feed was available for validation, it created its own testing system to verify that the code processed the data correctly.
This is an important difference between a normal chatbot and an AI agent. A chatbot may provide instructions for creating a test. An agent can recognize that validation is missing, build the test and use it to check its own implementation.
Adjustable Effort Settings
Opus 5 allows customers to adjust how much effort the model uses for a task.
A lower effort setting may be appropriate for simple questions, small code changes or routine document analysis. Higher effort settings allow the model to spend more time and tokens on difficult reasoning problems.
This gives developers and businesses more control over the relationship between:
- Response quality.
- Processing time.
- Token consumption.
- Total API cost.
Not every task requires the highest available reasoning setting. Using maximum effort for a simple formatting change may increase the cost without producing a meaningful improvement.
Claude Opus 5 Pricing
The standard API price for Claude Opus 5 is:
- Input: $5 per million tokens.
- Output: $25 per million tokens.
This is the same standard price as Opus 4.8.
Anthropic also offers a Fast mode that runs at approximately 2.5 times the normal speed. Fast mode costs twice the standard Opus 5 API price.
For businesses, the relevant cost will depend on more than the published token rate. A more capable model may cost more per token than a smaller model but complete a task with fewer retries, tool calls or manual corrections.
Developers should therefore measure the cost of completing the entire workflow rather than comparing only the price of individual tokens.
New API Model Name
Developers can access the model through Anthropic’s API using:
claude-opus-5Before changing an existing application, developers should test the new model with their own prompts, tool definitions and expected output formats.
A model that performs better in general benchmarks may still respond differently from an earlier version. Automated systems should be checked for changes in response length, tool selection, structured output and refusal behavior.
Changes to Tool Use and Automatic Fallbacks
Anthropic also introduced two related API features in beta.
The first allows developers to change which tools Claude can access during an active conversation without invalidating the existing prompt cache.
This could help applications provide tools only when they are needed, potentially reducing unnecessary access and improving security.
The second feature allows requests blocked by an Opus 5 or Fable 5 safety classifier to be automatically routed to another available model.
For example, Anthropic says requests flagged inside Claude, Claude Code or Claude Cowork can fall back to Opus 4.8 instead of ending immediately.
Automatic fallback may make applications more reliable, but developers should record which model actually completed each request. Different models may have different capabilities, costs and safety behavior.
What Anthropic Says About Safety
Anthropic says Opus 5 produced the lowest level of misaligned behavior among its recent models during the company’s automated behavioral audit.
The company also says it was less susceptible to deceptive instructions and attempts to misuse the model.
However, Opus 5 has become better at cybersecurity tasks as its general reasoning abilities have improved.
Anthropic says the model can identify software vulnerabilities at a level approaching its specialized Mythos 5 model, although it remains considerably less capable at turning those vulnerabilities into functioning exploits.
Opus 5 includes additional restrictions for penetration testing, binary vulnerability scanning and exploit generation. Verified cybersecurity professionals may obtain broader access through Anthropic’s Cyber Verification Program.
Opus 5 vs. Fable 5
Anthropic is not presenting Opus 5 as a complete replacement for Fable 5.
The company recommends Opus 5 for regular professional work where performance, speed and cost must remain balanced.
Fable 5 remains the more appropriate choice for extremely long and autonomous projects that may continue operating for days with limited human intervention.
A simplified comparison would be:
- Opus 5: coding, business analysis, research and everyday agentic work.
- Fable 5: the most demanding long-running autonomous projects.
- Mythos 5: specialized advanced cybersecurity and scientific capabilities.
The best model therefore depends on the task rather than simply choosing the most powerful option available.
Should Developers Switch Immediately?
Developers already using Opus 4.8 have a strong reason to test the new model because the standard API price has not increased.
However, changing the model used by a production application should still be treated as a software update.
Before completing a migration:
- Run the existing evaluation suite.
- Test important prompts and edge cases.
- Check tool calls and structured responses.
- Compare total token usage.
- Review safety refusals and fallback behavior.
- Verify latency under real workloads.
- Keep an easy way to return to the previous model.
Developers should also avoid relying entirely on benchmark results published by the model provider. Real project testing is the best way to determine whether the upgrade produces better results.
Why This Launch Matters
Claude Opus 5 reflects a broader shift in the artificial intelligence market.
AI laboratories are no longer competing only to release the model with the highest benchmark score. They are competing to provide models that can complete valuable work reliably and at a sustainable cost.
For developers, a slightly less powerful model may be more useful if it is faster, more predictable and inexpensive enough to use every day.
Anthropic’s strategy with Opus 5 is to bring performance close to its frontier models into a more accessible price range.
Final Thoughts
Claude Opus 5 appears to be one of Anthropic’s most practical model releases for developers and professional users.
It keeps the price of Opus 4.8 while promising better coding, research, computer use and long-running task performance.
Its ability to verify work and continue iterating may be especially valuable for coding agents, where an incomplete solution can be more harmful than an unsuccessful one.
However, the model should still be supervised, tested and evaluated against real workflows before being used for production decisions.
Claude Opus 5 is available now, and its success will depend less on benchmark charts than on whether it can help developers and businesses complete difficult work with fewer corrections and lower total costs.
Sources
- Anthropic’s official Claude Opus 5 announcement.
- Anthropic Claude Opus 5 System Card and product documentation.
- Reuters coverage published July 24, 2026.