Anthropic’s Opus 5 is about token efficiency, not a capability leap - Ars Technica

What the report says

Ars Technica reported that Anthropic has introduced Opus 5, an update to a model used for coding and software development work. The report characterizes the release less as a major capability jump and more as an efficiency and pricing move, with benchmark materials from Anthropic showing performance near or modestly above its Fable model on several coding-related tests, while improving incrementally over Opus 4.8 and competing systems cited in the article.

According to Ars Technica, the central selling point is that Opus 5 delivers performance close to Fable at roughly half the cost, rather than setting a new high-water mark for agentic coding. The article says Anthropic priced Opus 5 at $5 per million input tokens and $25 per million output tokens, matching its predecessor while undercutting Fable. Ars also notes that competition remains intense, citing the Chinese open-weight model Kimi K3 as being priced lower for output tokens while offering similar performance.

The report also highlights a cybersecurity distinction. Ars says Anthropic did not give Opus 5 the same advanced cybersecurity training as some other models, leaving it behind Fable and Mythos in exploitation-related tasks, even if Anthropic says it can help identify vulnerabilities. As a result, the model reportedly does not carry the same incident-review data retention policy associated with Fable.

The broader significance is cost pressure in AI software development. Ars points to growing interest in open-weight and local models, as well as “model routers” used by companies such as Cursor and Meta to send easier prompts to cheaper systems. That trend could push frontier-model providers to reduce token costs or offer more capability at the same price.