Anthropic’s Opus 5 focuses on cost efficiency rather than a major leap in capability

Anthropic has released Opus 5, the newest version of its popular coding-focused model. While the update delivers incremental benchmark improvements over Opus 4.8 and competitive models like OpenAI’s GPT-5.6-Sol, the headline story is about cost efficiency rather than a dramatic leap in capability.

According to benchmarks provided by Anthropic, Opus 5 performs at about the same level as the much more expensive Fable model on coding tasks like Frontier-Bench and DeepSWE. The key differentiator is price: Opus 5 sits at $5 per million input tokens and $25 per million output tokens, roughly half the cost of Fable.

This pricing comes as the developer community increasingly focuses on managing AI costs. Companies like Cursor and Meta are building “model routers” that automatically select the most cost-effective model for each task, reducing reliance on expensive frontier models for routine work. Meanwhile, open-weight alternatives like the Chinese Kimi K3 offer comparable performance at $15 per million output tokens, intensifying the competitive pressure.

Anthropic also made a deliberate choice to limit Opus 5’s cybersecurity training. While the model is relatively good at finding vulnerabilities, it lags significantly behind Fable and Mythos on exploitation tasks. This means Opus 5 does not include some of the controversial data retention policies that accompanied Fable’s release.

The broader trend here is that the AI model market is maturing. Capability gains are slowing, and the battleground is shifting to cost and efficiency. Anthropic’s strategy with Opus 5 appears to be a response to that shift: offer near-frontier performance at a price that keeps developers from migrating to cheaper alternatives.