NewYour coding agent can read the release notes before it upgrades.Set up the MCP server →
npm · #4234 most downloaded on npm
Last release today
25 Sep 2026
Ships on a steady schedule
a new release about every 8 days
Most releases are documented
notes for 38 of the last 60 stable releases
Nothing withdrawn
no release was ever pulled
2 years old
679 releases · first in 2024
Nothing published for this version
Nothing published for this version
### Patch Changes - Updated dependencies [31205a4] - @ai-sdk/provider-utils@4.0.46
One column per month.
### Patch Changes - Updated dependencies [31205a4] - @ai-sdk/provider-utils@4.0.46
Nothing published for this version
Nothing published for this version
7f803af : feat(alibaba): preserve reasoning in multi-turn requests by default on supported models
Nothing published for this version
af9597b : Compile packages for ES2022 runtime target
items arrays become prefixItems, type next to anyOf moves into the branches, and non-object root schemas fail with a clear client-side error instead of Moonshot's opaque 400. Everything else passes through unchanged; the original schema is still used for AI SDK result validation.### Patch Changes - Updated dependencies [31205a4] - @ai-sdk/provider-utils@4.0.46
Updated dependencies [ b5679a7 ]
@ai-sdk/openai-compatible; the converter, language model, and helpers are owned by the package. Video file parts (e.g. mediaType: 'video/mp4') are converted to Moonshot's video_url content parts for video-capable models (kimi-k3, kimi-k2.7-code, kimi-k2.6, kimi-k2.5). Audio and PDF file parts now throw client-side (the API rejects those part types). reasoningHistory: 'preserved' now maps to Moonshot's thinking.keep: 'all' request field (previously a no-op, the API ignores reasoning_history), gated per model with a warning on models without thinking.keep support. Adds promptCacheKey and safetyIdentifier provider options, widens reasoningEffort to 'low' | 'high' | 'max' per Moonshot's docs, maps the generic reasoning call option to reasoning_effort, and passes ms:// Files API references through natively (declared in supportedUrls).Nothing published for this version
Nothing published for this version
Updated dependencies [ 31d24ce ]
Patch Changes ai@7.0.110
Updated dependencies [ 0343bb1 ]
Nothing published for this version
a105059 : fix(workflow): support deferred tool discovery in WorkflowAgent
Updated dependencies [ 4775577 ]
Nothing published for this version
Nothing published for this version
Nothing published for this version
1f29230: feat(mcp): harden oauth client registration according to latest protocol
1011e33: Select MCP OAuth authorization scopes from WWW-Authenticate challenges or Protected Resource Metadata.
WWW-Authenticate challenges or Protected Resource Metadata.@ai-sdk/openai-compatible@3.0.31
Updated dependencies [ 615ac89 ]
Nothing published for this version
3b9f025: Add a provider-native reasoningEffort option for Open Responses endpoints.
reasoningEffort option for Open Responses endpoints.Nothing published for this version
Nothing published for this version
3204ab6 : fix(workflow): preserve per-tool types in execution callbacks and export the callback event types
Nothing published for this version
Updated dependencies [ 802af1e ]
180ce77 : fix(workflow): preserve tool behavior across workflow step boundaries
Nothing published for this version
Nothing published for this version
Nothing published for this version
Updated dependencies [ 850d863 ]
11109ae : feat(ai): support signed tool approvals in WorkflowAgent
Nothing published for this version
Updated dependencies [ 6669d69 ]
5a3ef3a: Fixed handling of image response in the tool call result.
### Patch Changes - Updated dependencies [99964ed] - @ai-sdk/provider-utils@3.0.8
a14fc2b: feat(provider/google): add gemini 2.5 flash image preview model support
### Patch Changes - Updated dependencies [886e7cd] - @ai-sdk/provider-utils@3.0.7
### Patch Changes - Updated dependencies [1b5a3d3] - @ai-sdk/provider-utils@3.0.6
### Patch Changes - Updated dependencies [8533108] - Updated dependencies [31205a4] - @ai-sdk/anthropic@3.0.111 - @ai-sdk/provider-utils@4.0.46
### Patch Changes - Updated dependencies [68751f9] - @ai-sdk/provider-utils@3.0.4
Updated dependencies [ 069a945 ]
Updated dependencies [ da2e17b ]
961dda1: add labels field to providerOptions
Updated dependencies [ f3c1980 ]
@ai-sdk/openai-compatible@2.0.69
@ai-sdk/openai-compatible@2.0.68
5604dde : feat(zai): add the Z.AI provider with GLM chat completions, streaming, reasoning, tools, and multimodal inputs
78e7fa9: Add code execution provider defined tool
f916255: feat (provider/google): add new gemini models
19a4336: Expose raw usageMetadata returned from Google Generative AI in providerMetadata
8af9e03: Added Image Models to the Google Provider for Imagen 3 Support
1a635b5: update supportedUrls to only support native URL
888b750: feat(providers/google): Add taskType support for Text Embedding Models
3259565: feat (providers/google): add thinking config to provider options
e2aceaf: feat: add raw chunk support
eb173f1: chore (providers): remove model shorthand deprecation warnings
6a16dcf: embed() now uses the single embeddings endpoint No code updates are needed.
This is to make sure that users are not ratelimited when using the batch endpoint, since many models have different limits for batch and single embeddings.
Eg: Google has a limit of 150 RPM for batch requests, and 1500 RPM for single requests.
Before, AI SDK would always use the batch endpoint, even for embed() calls, which led to ratelimits.
This does not have any breaking functionality and is fully tested :) if (values.length > 1) { const batchResult = await this.doEmbedBatch({ values, options, }); return batchResult; }
26735b5: chore(embedding-model): add v2 interface
5cf30ea: fix (provider/google): allow "OFF" for Google HarmBlockThreshold
443d8ec: feat(embedding-model-v2): add response body field
c68931f: Support tool schemas that allow additional properties (e.g z.record(z.string()))
66962ed: fix(packages): export node10 compatible types
a313780: fix: omit system message for gemma models
fd98925: chore(providers/google): update embedding model to use providerOptions
cb787ac: fix: remove non-functional models
7378473: chore(providers/google): switch to providerOptions
f07a6d4: fix(providers/google): accept nullish in safetyRatings
75f03b1: Add Gemini 2.5 Flash Lite GA
779d916: feat: add provider option schemas for vertex imagegen and google genai
581a9be: fix (provider/google): prevent error when thinking signature is used
2e06f14: feat (provider/google): Change to provider defined tools
8e6b69d: feat(providers/google): Add support for Gemini 2.5 Pro and Gemini 2.5 Flash (now stable)
42fcd32: feat(google): automatically handle system instructions for Gemma models
d1a034f: feature: using Zod 4 for internal stuff
fd65bc6: chore(embedding-model-v2): rename rawResponse to response
878bf45: removes (unsupported) additionalProperties from the Schema sent in the request payloads to Google APIs
0f05690: Add gemini-embedding-001 model, add embedding provider options type export
7badba2: fix(google): grounding streaming sources
205077b: fix: improve Zod compatibility
f10304b: feat(tool-calling): don't require the user to have to pass parameters
Updated dependencies
@ai-sdk/provider-utils@3.0.0-canary.19
@ai-sdk/provider-utils@3.0.0-canary.18
f07a6d4: fix(providers/google): accept nullish in safetyRatings
@ai-sdk/provider-utils@3.0.0-canary.16
Your coding agent can read these notes before it upgrades. Set up the MCP server →