Anthropic said on the 24th (local time) that it released Claude Opus 4.5, the latest version of Opus, the top model among its artificial intelligence (AI) models.
Anthropic offers three types of AI models—Opus, Sonnet, and Haiku—and Opus is a model suited for the most complex reasoning and professional tasks. Sonnet and Haiku are lighter-weight models. Anthropic said Opus 4.5 outperformed rival models, scoring 80.9% on the Software Engineering (SWE) benchmark that demonstrates agent coding capabilities.
OpenAI's GPT-5.1 Codex Max, built for engineering tasks, scored 77.9% on the same benchmark, and Google's "Gemini 3 Pro," released on the 18th, scored 76.2%. On the ARC-AGI-2 benchmark, which solves tasks not found in the training data, it also achieved 37.6%, outperforming Gemini 3 Pro (31.1%) and GPT-5.1 (17.6%).
Anthropic relayed feedback from customers who tried the new version early, saying that Opus 4.5 can now handle problems that even Sonnet 4.5 could not handle just a few weeks ago. Starting today, Opus 4.5 is available on the Claude app and API, as well as on the three major cloud platforms—Amazon Web Services (AWS), Microsoft Azure, and Google Cloud.
With the rollout of Opus 4.5, Anthropic also sharply cut API fees to $5 per 1 million input tokens and $25 per 1 million output tokens. That is one-third of the previous rates of $15 per 1 million input tokens and $75 per 1 million output tokens. One token is a unit into which AI splits text into arbitrary pieces for analysis. A single word may be one token, or it may be divided into multiple tokens.