Anthropic releases Opus 5, prioritising token efficiency over capability leaps
As the AI industry shifts focus toward cost management, Anthropic’s latest update highlights the growing role of model routers and the pressure from competitors like Kimi K3.
Anthropic has released Opus 5, an update to its popular coding and software development model, marking a strategic pivot toward token efficiency rather than significant capability breakthroughs. The release underscores a broader industry trend where cheaper model options are increasingly deemed sufficient for many tasks, reducing the necessity for the most expensive frontier models.
Priced at $5 per million input tokens and $25 per million output tokens, Opus 5 offers performance comparable to Anthropic’s Fable model at approximately half the cost. Benchmarks from Frontier-Bench and DeepSWE indicate that Opus 5 performs at a level similar to or slightly ahead of Fable on coding tasks. It also ostensibly outperforms its predecessor, Opus 4.8, and OpenAI’s GPT-5.6-Sol across various tasks, though the gains are described as iterative rather than radical.
Despite these performance metrics, the model lags behind in cybersecurity tasks due to specific training decisions. Anthropic deliberately excluded cutting-edge training on cybersecurity for Opus 5, resulting in performance that is “substantially behind Mythos 5” in exploiting vulnerabilities. While the company claims the model is “relatively good” at finding vulnerabilities, it does not include the controversial 30-day data retention policy for incident review that was present in Fable.
Competition is intensifying from cheaper alternatives, specifically the Chinese open-weight model Kimi K3, which offers similar performance at $15 per million output tokens. This pricing pressure highlights a shift in how software developers and engineering managers are approaching AI integration, with a growing focus on managing costs alongside performance.
To navigate this landscape, companies such as Cursor and Meta are increasingly adopting “model routers.” These systems automatically select cost-effective models based on prompt complexity, allowing users to save on compute and money by avoiding expensive frontier models for less challenging development tasks. Anthropic will need to continue reducing token costs or offering more performance for no additional cost to maintain its growth trajectory as users turn to smaller or open models.
