Repository navigation
Conversation
Raise floors to @ai-sdk/anthropic ^3.0.128, @ai-sdk/google-vertex ^4.0.211, and @ai-sdk/amazon-bedrock ^4.0.192. These pin @ai-sdk/provider 3.0.18 and provider-utils 4.0.57, so the rest of the AI SDK family moves with them to keep a single copy of the branded packages.
Haiku 5.5 has a 1M context window and 128k max output. Adaptive thinking
is on by default, but unlike Sonnet/Opus 5.5 it accepts
`thinking: {type: "disabled"}`, so the rule offers "off" alongside the
full effort range. The rule precedes the Haiku 4.5 rule.
Add the model to Anthropic's supplemental list until /v1/models returns
it, and cover Vertex display-name derivation for an undated Claude ID.
Snowflake Cortex does not list the model yet, so its catalog is unchanged.
"off" used to omit `thinking`, which on models whose default is adaptive
thinking (Haiku 5.5, Opus 5) left thinking on at the model's default
effort. "off" now sends `thinking: {type: "disabled"}` with no effort;
capability rules offer "off" only on models that accept it.
The five Claude-wire clients (Anthropic, Bedrock, Vertex, Snowflake,
Posit AI) each inlined the same adaptive-thinking options. They now share
`anthropicProviderOptions`, which maps undefined effort to no thinking
config, "off" to disabled, and any level to summarized adaptive thinking,
and merges client options such as `toolStreaming: false` on top.
Final-wire tests run the real SDK serializer for Haiku 5.5, Opus 5, and
Opus 4.8; routing tests cover the other four clients.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
This PR adds Claude Haiku 5.5 to model discovery and capability inference so applications can offer its 1M-token context window, 128K-token output limit, and supported thinking levels. It also makes Claude's “Off” thinking setting disable thinking on models that support it, instead of leave adaptive thinking enabled by default.
The provider library updates the Anthropic, Vertex, and Bedrock SDK minimum versions together and uses one thinking-options implementation across Anthropic, Bedrock, Google Vertex, Snowflake, and Posit AI Pass. Active effort levels request adaptive thinking with visible summaries; provider-default requests leave the thinking mode and effort unspecified. Existing route-specific tool-streaming options remain intact.
Haiku 5.5 offers “off,” “low,” “medium,” “high,” “xhigh,” and “max.” “Off” sends an explicit disable request without an effort value, avoiding the incompatible combination of disabled thinking and the highest effort levels. Claude models that cannot disable thinking still do not offer “Off.” Model metadata recognizes bare IDs and supported provider prefixes, and Vertex displays the undated model as “Claude Haiku 5.5.”
Availability still depends on the provider. Snowflake's catalog is unchanged because Cortex does not list Haiku 5.5; the Posit AI Pass gateway is not assumed to serve it merely because the capability rule exists.
Verification
Regression coverage checks capability inference for provider-prefixed IDs, Vertex display names, thinking options on all five Claude routes, and actual Anthropic SDK request serialization for provider-default, disabled, and active thinking. Implementation verification recorded in the companion plan also covers live Anthropic and Bedrock requests; live Vertex verification remains untested because Google credentials were unavailable.