Claude Sonnet 5.5 Is Here — and Anthropic Quietly Admits Your Prompt Might Not Get the Model You Picked

Claude Sonnet 5.5 has been released on September 28, six days after Opus 5.5. This top story on Hacker News garnered 775 points and over 500 comments. However, buried under benchmark charts is a concerning detail: Anthropic may quietly swap your request to the older Sonnet 5 if they deem your prompt to be higher-risk. Here are the details: The pricing remains the same at $2/M input and $10/M output.

However, performance has improved significantly. For terminal-agentic coding, Sonnet 5.5 scores 70.6%, a substantial jump from its benchmark of 10.3% in Sonnet 5. It outperforms Opus 5.5 in general knowledge work with a score of 1844, just two points shy of Opus 5.5's 1846 and significantly above Sonnet 5's 1449. In chartography, Sonnet 5.5 scores 61.6%, a marked improvement over Sonnet 5's 15.6%.

Anthropic claims that Sonnet 5.5 is 30% faster and up to 30% cheaper due to fewer tokens and tool calls. Devs have reported that they have returned to Sonnet 5.5 due to lower token consumption. However, there's a catch. Sonnet 5.5 comes with "frontier-style cyber safeguards" capable of detecting and blocking misuse patterns. If Anthropic classifies your request as "higher-risk," they can silently route it to the older Sonnet 5 model.

This means you may receive responses from a different model than the one you requested, which could be problematic for deterministic model behavior. For straightforward coding or general knowledge tasks, switching to Sonnet 5.5 is recommended. However, for workloads depending on consistent model identity, it's advised to investigate what triggers this "higher-risk" reroute and log the model field in every response.