- Sources: primary, discussion
- Summary: Integrations that ran Sonnet with thinking disabled have to switch to the new
between_tools setting before moving to Sonnet 5.5, and higher-risk cybersecurity requests visibly fall back to Sonnet 5 under the model's safeguards. Anthropic's own page states the model does not advance the frontier of its capabilities. Every benchmark figure on that page is Anthropic's own, the GDPval-AA v2.1 and AA-Briefcase v1.1 results were run by Artificial Analysis against a pre-release deployment that Anthropic says carried a structured-outputs bug since fixed, and several comparison rows substitute GPT-5.6 Sol because GPT-6 Sol figures were not published for Terminal-Bench 4.0 or CursorBench 4.0. - Why it matters: This is a configuration change and a behaviour change rather than a drop-in swap, so an integration that only updates a model string inherits both.
send feedback on this story