Skip to main content

AI SDK Dependency Upgrade Research

Date: 2026-02-27 Researcher: ai-sdk-researcher (automated) Scope: 7 AI SDK packages from vercel/ai monorepo

Executive Summary​

These upgrades are low risk overall. The most significant changes are:

  1. Security fix in @ai-sdk/provider-utils 4.0.15: download size limits to prevent memory exhaustion (DoS)
  2. New feature in @ai-sdk/openai 3.0.36: fix for Azure AI Foundry/Mistral streaming tool calls
  3. New feature in @ai-sdk/openai 3.0.35: enhanced reasoning content (Responses API)
  4. New feature in @ai-sdk/openai 3.0.34: phase parameter support for gpt-5.3-codex
  5. New feature in @ai-sdk/anthropic 3.0.48: code execution tool support
  6. New feature in @ai-sdk/google 3.0.32-3.0.33: Gemini 3.1 image model support
  7. Bug fix in ai 6.0.101: duplicate tool part creation for non-existent tools

No breaking changes were found in any of these upgrades.

1. @ai-sdk/anthropic (3.0.47 -> 3.0.48)​

What Changed​

VersionTypeDescription
3.0.48FeatureAdded support for a new code execution tool (2164cdf)
3.0.47FixChanged cacheControl provider option to pass as top-level cache_control parameter in Anthropic API request body, enabling automatic caching (17978c6)

Breaking Changes​

None.

New Features We Can Leverage​

  • Code execution tool: Anthropic's code execution (sandbox) capability is now supported. NeuroLink could expose this as a built-in tool option for Anthropic provider users, similar to how we handle MCP tools.
  • Improved cacheControl: The cache_control parameter is now placed correctly at the top level of the API request, which means prompt caching will work more reliably. NeuroLink's Anthropic provider should benefit automatically.

Risk Level​

Low - Patch-level changes with additive features only. The cacheControl change was already in 3.0.47 (current version).

Security Fixes​

None in these versions directly, but the transitive dependency update to @ai-sdk/provider-utils carries a security fix (see provider-utils section below).

2. @ai-sdk/azure (3.0.35 -> 3.0.37)​

What Changed​

VersionTypeDescription
3.0.37DepsUpdated dependency: @ai-sdk/[email protected]
3.0.36DepsUpdated dependency: @ai-sdk/[email protected]
3.0.35DepsUpdated dependency: @ai-sdk/[email protected]

Breaking Changes​

None.

New Features We Can Leverage​

All features come transitively from @ai-sdk/openai updates (see section 6). Most notably:

  • Streaming tool call fix (3.0.36 via [email protected]): Azure AI Foundry deployments that omit the type field in streaming tool_calls deltas no longer throw InvalidResponseDataError. This is a direct fix for Azure users of NeuroLink.
  • Reasoning content fallback (via [email protected]): Multi-turn reasoning works even when item IDs are stripped.
  • Phase parameter (via [email protected]): Support for gpt-5.3-codex phase field.

Risk Level​

Low - Pure dependency bumps. The Azure package itself has no code changes.

Security Fixes​

None directly, but inherits the download size limit fix from provider-utils.

3. @ai-sdk/google (3.0.31 -> 3.0.33)​

What Changed​

VersionTypeDescription
3.0.33FeatureAdded support for new Google image model aspect ratios and sizes (1ece97a)
3.0.32FeatureAdded compatibility for gemini-3.1-flash-image-preview model (45f0a7f)
3.0.31TypesExpanded GoogleGenerativeAIModelId and GoogleGenerativeAIVideoModelId type definitions for better autocomplete

Breaking Changes​

None.

New Features We Can Leverage​

  • Gemini 3.1 Flash Image Preview model: NeuroLink's Google AI Studio provider can now use the gemini-3.1-flash-image-preview model for image generation tasks. Consider adding this to model definitions.
  • Image aspect ratios/sizes: Users can specify more granular image output dimensions. NeuroLink's image generation API should pass through these options.
  • Better type definitions: Improved autocomplete for model IDs. No action needed - automatic benefit.

Risk Level​

Low - Additive features only, no behavior changes to existing functionality.

Security Fixes​

None.

4. @ai-sdk/google-vertex (4.0.63 -> 4.0.66)​

What Changed​

VersionTypeDescription
4.0.66DepsUpdated @ai-sdk/[email protected]
4.0.65DepsUpdated @ai-sdk/[email protected]
4.0.64Feature + DepsAdded support for gemini-3.1-flash-image-preview model; updated @ai-sdk/[email protected]
4.0.63DepsUpdated @ai-sdk/[email protected]

Breaking Changes​

None.

New Features We Can Leverage​

  • Gemini 3.1 Flash Image Preview on Vertex: Same model support as @ai-sdk/google but through Google Vertex AI. NeuroLink's Google Vertex provider gets this automatically.
  • Anthropic on Vertex: Gets code execution tool support via the anthropic dependency bump.

Risk Level​

Low - Primarily dependency updates. One additive model feature.

Security Fixes​

None directly.

5. @ai-sdk/mistral (3.0.12 -> 3.0.20)​

What Changed​

This is the largest version jump (8 versions), but almost entirely dependency and documentation updates.

VersionTypeDescription
3.0.20DepsUpdated @ai-sdk/[email protected]
3.0.19DepsUpdated @ai-sdk/[email protected], @ai-sdk/[email protected]
3.0.18DepsUpdated @ai-sdk/[email protected], @ai-sdk/[email protected]
3.0.17DepsUpdated @ai-sdk/[email protected]
3.0.16DepsUpdated @ai-sdk/[email protected], @ai-sdk/[email protected]
3.0.15DocsAdded skill information to README files
3.0.14DocsFixed incorrect and outdated provider docs
3.0.13DepsUpdated @ai-sdk/[email protected]
3.0.12HousekeepingExcluded tests from npm package; dependency updates

Breaking Changes​

None.

New Features We Can Leverage (via transitive dependencies)​

  • Download size limits ([email protected]): Security fix - prevents memory exhaustion from oversized downloads.
  • Video model resolution ([email protected]): Default global provider video model resolution.
  • Experimental video support ([email protected]): Experimental generateVideo support added to provider interface.
  • Better error messages ([email protected], [email protected]): Type validation errors now include field paths and entity identifiers.
  • Bun compatibility ([email protected]): Bun fetch errors are now recognized as retryable.
  • Type export fix ([email protected]): Only exports types from standard-schema package.

Risk Level​

Low - No Mistral-specific code changes. All changes are in shared dependencies.

Security Fixes​

Yes - Transitive via @ai-sdk/[email protected]:

  • Download size limit enforcement (default 2 GiB max) to prevent memory exhaustion DoS
  • abortSignal now properly passed to fetch() across all download call sites

6. @ai-sdk/openai (3.0.34 -> 3.0.36)​

What Changed​

VersionTypeDescription
3.0.36Bug FixFixed streaming tool call handling to accept null/undefined type fields. Azure AI Foundry and Mistral deployments on Azure omit the type field in streaming tool_calls deltas, which previously caused InvalidResponseDataError. Parser now treats missing type as "function" instead of failing.
3.0.35EnhancementEnhanced reasoning content part handling in the Responses API. When providerOptions.openai.itemId is absent on reasoning content parts, the converter now uses encrypted_content as a fallback instead of skipping the part. Made id field optional on OpenAIResponsesReasoning type.
3.0.34FeatureAdded support for the phase parameter on Responses API message items. Models like gpt-5.3-codex return phase fields ('commentary' or 'final_answer') on assistant message output items. Values preserved in providerMetadata.openai.phase on text parts.

Breaking Changes​

None.

New Features We Can Leverage​

  • Streaming tool call fix (3.0.36): This is a critical fix for NeuroLink's Azure provider. Users deploying Mistral models on Azure AI Foundry will no longer get InvalidResponseDataError during streaming tool calls. This was likely causing failures for NeuroLink users.
  • Reasoning content fallback (3.0.35): Multi-turn conversations with reasoning models work better. NeuroLink's OpenAI provider benefits automatically when using the Responses API.
  • Phase parameter (3.0.34): Support for gpt-5.3-codex model's phase field. NeuroLink could expose phase metadata in its response objects. Important: correctly preserving phase on assistant items is required for gpt-5.3-codex - dropping it causes significant performance degradation.

Risk Level​

Low - All changes are backward-compatible. The streaming fix (3.0.36) actually resolves existing failures.

Security Fixes​

None directly.

7. ai (6.0.101 -> 6.0.103)​

What Changed​

VersionTypeDescription
6.0.103DepsUpdated @ai-sdk/[email protected]
6.0.102DepsUpdated @ai-sdk/[email protected]
6.0.101Bug FixFixed duplicate tool part creation when models invoke non-existent tools (5230482)

Breaking Changes​

None.

New Features We Can Leverage​

  • Duplicate tool part fix (6.0.101): When a model hallucinates a tool name that doesn't exist, the SDK no longer creates duplicate tool parts. This improves reliability of NeuroLink's tool execution pipeline, especially with less capable models that may hallucinate tool names.

Risk Level​

Low - Bug fix and dependency bumps only.

Security Fixes​

None directly, but the gateway dependency updates may carry transitive fixes.

Transitive Dependency Changes (Important)​

@ai-sdk/provider-utils (4.0.9 -> 4.0.15)​

This is the most significant transitive dependency and carries a security fix:

VersionTypeDescription
4.0.15SECURITYdownload() and downloadBlob() now enforce a default 2 GiB size limit on user-provided URLs. Downloads exceeding the limit abort with DownloadError. abortSignal properly passed to fetch(). New createDownload({ maxBytes }) factory.
4.0.14DepsUpdated @ai-sdk/[email protected]
4.0.13DepsUpdated @ai-sdk/[email protected]
4.0.12FixExport only types from standard-schema package (removes import conflicts)
4.0.11EnhancementType validation error messages include field paths and entity identifiers
4.0.10FixRecognize Bun fetch errors as retryable
4.0.9HousekeepingExcluded tests from npm package

@ai-sdk/provider (3.0.5 -> 3.0.8)​

VersionTypeDescription
3.0.8FeatureDefault global provider video model resolution
3.0.7FeatureExperimental generate video support
3.0.6FixType validation error messages include field paths and entity identifiers
3.0.5HousekeepingExcluded tests from npm package

Known Security Advisories​

CVE-2025-48985: Input Validation Bypass (AI SDK)​

  • Severity: Low
  • Affected versions: AI SDK < 5.0.52 and 6.0.0-beta.*
  • Description: Improper URL-to-data mapping allows attackers to substitute arbitrary downloaded bytes for different supported URLs within the same prompt. Filtering operations cause index misalignment between downloaded files and their intended URLs.
  • Status: Fixed in versions we are already past (we are on 6.0.101+). Not a concern for this upgrade.
  • Affected functions: generateText(), streamText(), and most methods accepting images/files as input.

Download Size Limit (provider-utils 4.0.15)​

  • Severity: Medium (DoS prevention)
  • Description: Prior to 4.0.15, download() and downloadBlob() had no size limit, allowing potential memory exhaustion when processing user-provided URLs.
  • Status: Fixed in @ai-sdk/[email protected], which is pulled in by @ai-sdk/[email protected] and will be transitively pulled in by all provider packages.
  • Impact on NeuroLink: If NeuroLink passes user-provided URLs to generateText()/streamText() (e.g., image URLs), this fix prevents a potential DoS vector.

Upgrade Recommendations​

Priority Order​

  1. @ai-sdk/openai 3.0.36 + @ai-sdk/azure 3.0.37 - Fixes Azure streaming tool call failures
  2. @ai-sdk/mistral 3.0.20 - Brings in the security fix for download size limits
  3. ai 6.0.103 - Bug fix for duplicate tool parts
  4. @ai-sdk/anthropic 3.0.48 - Code execution tool support
  5. @ai-sdk/google 3.0.33 - New image model support
  6. @ai-sdk/google-vertex 4.0.66 - Dependency alignment

Overall Risk Assessment: LOW​

All 7 packages are safe to upgrade simultaneously:

  • Zero breaking changes
  • All semver-compliant patch updates
  • One security-relevant fix (download size limits)
  • One important bug fix (Azure streaming tool calls)
  • Several additive features (code execution, image models, phase parameter)
  1. After upgrading, verify Azure provider streaming with tool calls works correctly
  2. Consider exposing Anthropic code execution tool in NeuroLink's tool system
  3. Consider adding gemini-3.1-flash-image-preview to model definitions
  4. Consider preserving phase metadata from gpt-5.3-codex responses
  5. Ensure NeuroLink passes through the download size limit options if users need to customize the 2 GiB default