Hotfix for Home Assistant 2026.11 and Claude Opus 4.7. Bundled work merged since 3.45.0 ships in 3.46.0.
Fixed
- Claude Opus 4.7 works with Thinking on (#932). Claude Opus 4.7 is offered for every Anthropic feature, but it no longer accepts the extended-thinking request HGA sent (
budget_tokens), so with Thinking on every request to it failed. It now gets adaptive thinking ({"type": "adaptive"}) with an effort level read from the thinking budget (up to 2048low, up to 8192medium, otherwisehigh). Opus 4.7 also rejectstemperature,top_pandtop_kon any API version, so they are no longer sent to it, with olderanthropicSDKs too (Home Assistant before 2026.11). The same applies to later Claude models. Sonnet 4.6 and Haiku 4.5 keep extended thinking with a budget. - Home Assistant 2026.11: the Anthropic provider works next to Home Assistant's own Anthropic integration (#929). Home Assistant 2026.11's Anthropic integration installs the
anthropic2.x SDK, which thelangchain-anthropicrelease HGA allowed (1.4.3) cannot run with, and which talks to a newer Anthropic API version (2026-10-21). After upgradinglangchain-anthropicby hand, every chat, camera, summarization and Sentinel request to an Anthropic model failed withtemperature: Extra inputs are not permitted, and with a smaller model such as Claude Haiku 4.5 then withmax_tokens: 128000 > 64000. HGA now requireslangchain-anthropic1.7.10 or later, which runs with the 2.x SDK (tested withanthropic2.3.0;langchain-anthropicitself only promises 0.x and 1.x). Under that SDK it no longer sendstemperature,top_portop_kto Anthropic, which that API version does not accept, so the Temperature setting has no effect on Anthropic models there; with an older SDK it is sent as before. A request that does not sizemax_tokensitself now asks for 16,384 tokens, or the model's own limit if that is lower, instead of the output limit of Claude Sonnet 4.6 for every model; extended thinking still sizes it from the thinking budget.