Vercel AI SDK Adds Support for Grok 4.6 Model IDs and 'xhigh' Reasoning Effort

Vercel AI SDK Adds Grok 4.6 Model IDs and 'xhigh' Reasoning Effort

Vercel shipped version 4.0.37 of the @ai-sdk/xai package, a patch release that adds support for Grok 4.6 model IDs and introduces a new "xhigh" reasoning effort setting for xAI's models.

The change, tracked in commit a4d386d, is described in the release notes as: "feat(xai): add the Grok 4.6 model IDs and support its xhigh reasoning effort." The release was published automatically by GitHub Actions on August 12 at 15:54 UTC, one commit ahead of the main branch at time of tagging.

What changed

This is a patch-level update (4.0.37) to the xAI provider package within the Vercel AI SDK monorepo. Patch releases in this package typically add compatibility for new upstream model identifiers or parameters without altering the SDK's public API surface. In this case, the update does two things:

• Registers new model ID strings for Grok 4.6, allowing developers to reference the model directly in ai-sdk calls (e.g., via xai('grok-4.6')-style identifiers).
• Adds support for an "xhigh" reasoning effort tier, extending the existing reasoning-effort parameter options that xAI exposes for its Grok reasoning models.

No new benchmark data, pricing, or context window figures accompanied this SDK release. The underlying Grok 4.6 model itself was not detailed in this changelog — this release only concerns the SDK's ability to address and configure it. Specific parameters for Grok 4.6, including context window size, input/output pricing per million tokens, and benchmark scores, have not been disclosed in this release and were not part of xAI's own announcement channels referenced here.

Context

The @ai-sdk/xai package is one of many provider integrations maintained inside Vercel's ai monorepo, which lets developers write model-agnostic code across OpenAI, Anthropic, xAI, Google, and other providers through a unified interface. Reasoning-effort settings — low, medium, high, and now xhigh — let developers trade off latency and cost against how much internal deliberation a reasoning model performs before returning an answer, a pattern xAI has used since earlier Grok reasoning releases.

What this means

This release is infrastructure plumbing, not a model launch. It confirms that xAI has model IDs for something called "Grok 4.6" and that it supports a new, more intensive reasoning tier called "xhigh," but xAI has not published its own detailed model card, pricing, or benchmarks alongside this SDK update. Developers building on the Vercel AI SDK can now target Grok 4.6 in code, but anyone evaluating the model's actual capabilities, cost, and context window will need to wait for xAI's own release documentation.