Anthropic’s Opus 5 Prioritizes Token Efficiency Over Capability
▼ Summary
– Anthropic released Opus 5, the newest update to its popular coding model.
– Opus 5 performs at about the same level or slightly ahead of Anthropic’s Fable model on coding benchmarks, but is not a breakthrough like Opus 4.5.
– The model offers performance just shy of Fable at approximately half the cost.
– Opus 5 lags behind Fable and Mythos in cybersecurity tasks, especially in exploiting vulnerabilities.
– Unlike Fable, Opus 5 does not include controversial data retention policies, such as keeping data for 30 days.
Anthropic has officially launched Opus 5, the latest iteration of its popular model known for excelling in coding and software development workflows. The release marks a meaningful update, though it does not represent the kind of transformative leap seen with the Opus 4.5 generation, particularly in agentic coding performance.
According to a benchmark chart produced by Anthropic, Opus 5 delivers results that are roughly on par with, or slightly ahead of, the company’s highly touted Fable model for programming tasks. Across evaluations like Frontier-Bench and DeepSWE, Opus 5 consistently outperforms both Opus 4.8 and OpenAI’s competing GPT-5.6-Sol in nearly every category.
While these benchmarks show steady, iterative gains rather than a radical breakthrough, the primary selling point is cost efficiency. Anthropic positions Opus 5 as a model that delivers performance just shy of Fable at roughly half the price, making it a compelling option for budget-conscious developers.
Notably, Anthropic deliberately chose not to equip Opus 5 with cutting-edge training in cybersecurity. As a result, the model lags significantly behind both Fable and Mythos in this domain. While Anthropic claims Opus 5 is reasonably adept at identifying cybersecurity vulnerabilities, training decisions mean it is “substantially behind Mythos 5 on the exploitation of those vulnerabilities.”
This strategic trade-off also means Opus 5 lacks some of the controversial safeguards that defined Fable’s launch, such as the policy requiring a 30-day data retention window for incident review. The result is a model that prioritizes token efficiency and cost savings over raw capability, especially in high-stakes security contexts.
(Source: Ars Technica)
