Opus 5 from Anthropic: Focus on Token Efficiency, Not a Leap in Capabilities
Anthropic
OpenAI
Anthropic has released Opus 5, a new version of its model. The main emphasis is on cost efficiency: the model performs almost like the more powerful Fable but at roughly half the price. On benchmarks, Opus 5 shows iterative improvement, but lags in cybersecurity tasks due to training specifics.
Anthropic has introduced Opus 5, the latest update to its model popular for programming and development. The release is not a breakthrough on the level of Opus 4.5 in agentic coding. According to Anthropic's charts, Opus 5 achieves results comparable to the Fable model in coding tasks and outperforms Opus 4.8 and OpenAI GPT-5.6-Sol on nearly all tests, but the performance improvement is iterative rather than radical. The main advantage is cost: Opus 5 offers capabilities close to Fable for roughly half the price. Meanwhile, Anthropic deliberately did not train Opus 5 on advanced cybersecurity tasks, so the model significantly lags behind Fable and Mythos in this area. Opus 5 is decent at finding vulnerabilities but falls substantially short of Mythos 5 in exploiting them. Unlike Fable, Opus 5 does not have the same strict safeguards, such as 30-day data retention for incident analysis.
Source: Ars Technica —
original