Anthropic launches Claude Opus 4.5 for agents and coding

Anthropic has launched Claude Opus 4.5, an artificial intelligence model designed for programming, autonomous agents, and tasks that require using tools over long periods. It is now available in Claude apps, through its API, and on the three major cloud platforms.

The company presents it as its most capable model yet for software engineering and computer use. It also says it performs better in research, mathematics, vision, spreadsheets, and presentations.

More capability at a lower cost

The most relevant change is not just that Opus 4.5 can solve difficult problems. According to Anthropic, it also needs fewer steps and generates less text to reach an answer, which can reduce costs when it is used at scale.

The API price is $5 per million input tokens and $25 per million output tokens. Tokens are the units of text the model processes. For developers, the access identifier is claude-opus-4-5-20251101.

Anthropic has added a parameter called effort, which lets you decide how much the model should reason before responding. At a medium level, Opus 4.5 matched Sonnet 4.5's best score on SWE-bench Verified, a programming test, while using 76% fewer output tokens. At the highest level, it outperformed Sonnet 4.5 by 4.3 percentage points while using 48% fewer tokens.

The idea is simple: not every task needs the same level of reasoning. For a small code change, you can prioritize speed and cost. To migrate an entire system, you can let the model spend more time planning.

Designed for long-running, autonomous tasks

Opus 4.5 is designed for agents, systems that do more than respond and instead execute a chain of actions using external tools. It can search for information, edit files, run tests, and fix errors without asking for instructions at every step.

In Terminal Bench, an evaluation of tasks performed from the command line, Anthropic reports a 15% improvement over Sonnet 4.5. In another deep research test, the combination of memory, context management, and multiple sub-agents raised the result from 70.48% to 85.30%.

The company also says the model used up to 65% fewer tokens on long-running programming tasks while maintaining better success rates on held-out tests. These results come from Anthropic evaluations and customers with early access, not from a single independent measurement.

One of the most notable figures comes from an internal technical test for performance engineering candidates. Within the two-hour limit, Opus 4.5 scored higher than any previous human candidate. Anthropic clarifies that the test measures technical ability and judgment under pressure, but not skills such as communication, collaboration, or professional experience.

It is also coming to workplace tools

The launch includes changes to several Anthropic products:

• Claude Code now includes a more precise planning mode. The system asks questions before starting and creates an editable file called plan.md.
• The desktop app can run several Claude Code sessions in parallel, for example, one to fix bugs, another to investigate a repository, and another to update documentation.
• Long conversations in Claude no longer stop when they reach the context limit. The system automatically summarizes earlier parts.
• Claude for Chrome is now available to all Max users.
• Claude for Excel is expanding its beta to Max, Team, and Enterprise users.

For you, this could mean something concrete: asking Claude to review a project, consult documentation, modify several files, and check the result with less manual intervention. In Excel, it can help build financial models or automate repetitive tasks, although its output still needs review when money or important decisions are involved.

Harder to trick, but not infallible

Anthropic says Opus 4.5 is its most resilient model yet against prompt injection attacks. This type of attack attempts to insert hidden instructions into a website, file, or email to divert the model from its original task.

The model also showed a useful but risky capability: finding solutions that were not anticipated by a rule. In an air travel test, it discovered that it could first change the ticket class and then modify the flight, even though the basic fare did not allow direct changes. The evaluation system marked this as a failure because the strategy was not the expected one.

That example sums up the challenge of AI agents. Solving more problems independently is valuable, but it also makes it more important to set clear limits, review their actions, and prevent them from optimizing a goal in ways no one anticipated.

Claude Opus 4.5 brings the most capable models closer to everyday work tasks, not just programming demonstrations. What matters now is whether its improvements in planning and autonomy hold up outside controlled tests, especially when it handles files, accounts, and decisions with real-world consequences.