Anthropic's Opus 5 matches Fable's coding performance at half the cost

Today, Anthropic released Opus 5, the newest version of its coding-focused model. Benchmarks show it performs on par with or slightly ahead of Anthropic's top-tier Fable model for software development tasks, while beating previous Opus 4.8 and OpenAI's GPT-5.6-Sol across nearly all evaluated tasks. However, this is not a radical capability leap like the jump from Opus 4 to Opus 4.5. Instead, the primary advantage is token efficiency: Opus 5 costs $5 per million input tokens and $25 per million output tokens, roughly half the price of Fable. Anthropic deliberately avoided training Opus 5 on advanced cybersecurity tasks, so it lags behind Fable and Mythos in vulnerability exploitation, though it remains capable at finding basic vulnerabilities.

The release highlights a broader industry shift toward cost optimization. Many developers and engineering managers are now focusing on managing token expenses, prompting companies like Cursor and Meta to build "model routers" that automatically select the most appropriate model size for each prompt, saving money by reserving high-end models like Fable for complex tasks. Anthropic must continue reducing token costs to compete with emerging open-weight models such as Kimi K3, which offers similar performance at just $15 per million output tokens. Opus 5 maintains the same pricing as its predecessor but delivers more performance per dollar. Without such efficiency gains, usage could shift to smaller or open models once they become adequate for routine development work. This strategic move aligns with Anthropic's goal of keeping enterprise customers despite fierce competition.