Anthropic launches Claude Opus 4.5 and cuts price by two-thirds
Claude Opus 4.5 leads in software engineering and long-running agent tests. Anthropic has cut its price to $5 per million input tokens and $25 per million output tokens, bringing its most capable model closer to everyday use.
Anthropic today unveiled Claude Opus 4.5, the latest version of its most capable model. The company says it has reclaimed the top spot on SWE-bench Verified, one of the most closely watched tests of a model’s ability to resolve real-world programming issues, and improved at tasks that require it to work autonomously for long periods.
The change is not merely technical. Anthropic has cut Opus pricing to $5 per million input tokens and $25 per million output tokens, down from $15 and $75 for the previous generation. That is a nearly 66% reduction, changing Opus’s role: it is no longer reserved for exceptionally costly workloads and is now competing to become the go-to model for development teams.
The race for the flagship model accelerates
The launch caps an unusual concentration of announcements. OpenAI updated GPT-5 with GPT-5.1 on November 12; Google introduced Gemini 3 on November 18; and Anthropic is now responding with Opus 4.5. All three major providers have refreshed their flagship model in just 12 days.
That pace reflects the fact that competition is no longer decided solely by which model answers questions or writes text better. The battlefield is agents: systems that receive a goal, use tools such as a terminal or browser, and chain together steps to complete a task. In programming, that means exploring a repository, locating a bug, changing multiple files, running tests, and fixing any errors that emerge.
Anthropic positions Opus 4.5 as its best model for coding, computer use, and agents. It also touts advances in deep research, spreadsheets, presentations, vision, mathematics, and reasoning. Its apps, API, and the three major cloud platforms are making it available today under the identifier claude-opus-4-5-20251101.
More performance with fewer steps
The core promise of Opus 4.5 is not just greater accuracy, but reaching the result with fewer tool calls and fewer tokens. That matters because an agent’s cost depends on more than the model’s rate: a task that gets stuck, repeats searches, or runs too many checks can multiply the bill and take far longer than expected.
Among the evaluations Anthropic cites, the model outperforms Sonnet 4.5 on Aider Polyglot — a programming test spanning multiple languages — and delivers 29% better performance on Vending-Bench, which is designed to test whether an agent stays on track during extended tasks. The company also says it leads in seven of the eight languages evaluated in SWE-bench Multilingual.
The company also says Opus 4.5 scored higher, in an internal performance-engineering exam with a two-hour limit, than any human candidate who had taken it. It is a striking comparison, but it needs to be put in context: the test measures technical skills and judgment under pressure in a specific setting, not professional abilities such as collaborating with a team, communicating decisions, or understanding a company’s context.
Benchmarks have limits too
Anthropic includes a revealing example from a test called τ2-bench. In an airline customer-service scenario, the agent had to help change a flight with a basic fare that does not allow modifications. Opus 4.5 found a route permitted by the rules: first change the cabin class, then modify the flight. The test marked it as an error because that solution was not included in the expected answer.
The case highlights a growing tension in agent evaluation. A benchmark makes it possible to compare models repeatedly, but it can penalize a correct solution if it is too rigid. At the same time, the ability to find unanticipated paths requires better-designed rules and oversight: in real-world environments, an ingenious solution must not bypass important constraints.
For developers and businesses, the most tangible news is the combination of capability and price. Opus 4.5 could prove attractive for refactoring code, migrating legacy systems, reviewing complex changes, or automating processes that use multiple tools. Even so, a lead on public tests does not eliminate the need to review generated code, control permissions, and measure costs in specific use cases. The decisive test will be whether that improvement holds up beyond benchmarks, in projects involving real-world data, rules, and errors.
