Vercel AI SDK Patch Adds Support for 'None' Reasoning Effort on GPT-6 Sol and Luna

Vercel shipped version 4.0.78 of the @ai-sdk/openai package on September 26, extending support for the reasoningEffort parameter to accept a value of 'none' when used with GPT-6 Sol and GPT-6 Luna models.

The patch, tracked under commit 8ff139b, modifies how the SDK handles reasoning-effort configuration at both the request level and in positioned system messages. According to the release notes, the update validates any effort change against what a given model actually supports, rather than passing values through unchecked.

What changed

The changelog describes two specific behaviors introduced in this patch:

• Request-level updates: If a developer sets an unsupported reasoning-effort value in a request, the SDK now logs a warning and omits the invalid update rather than sending it to the API.
• Historical updates: If an unsupported effort level appears in message history (for example, a positioned system message), the SDK rejects it outright instead of silently passing it through.

This is a patch-level change (4.0.77 → 4.0.78), meaning it does not alter the package's public API surface in a breaking way — it refines validation logic around an existing parameter.

The GPT-6 Sol and Luna mention

The release notes name two model identifiers not previously documented in OpenAI's public model lineup: GPT-6 Sol and GPT-6 Luna. Their appearance in an SDK patch changelog — rather than in an OpenAI announcement — suggests these are either internal codenames, limited-access models, or models in a private beta that Vercel's SDK team has already begun wiring up support for. OpenAI has not published pricing, context window specifications, benchmark scores, or a public release date for either model as of this writing. Nothing here should be read as confirmation of a public GPT-6 launch; it is only confirmation that the AI SDK's OpenAI provider now recognizes these model names in its reasoning-effort validation logic.

Why reasoning effort matters

The reasoningEffort parameter, used across OpenAI's o-series and newer reasoning-capable models, lets developers trade off latency and cost against depth of chain-of-thought computation. Adding a 'none' option implies these two models support a mode where reasoning is disabled entirely, likely for use cases prioritizing speed or cost over deliberation.

What this means

This is a routine SDK maintenance release, not a model launch — no new model card is warranted here. But the naming of "GPT-6 Sol" and "GPT-6 Luna" inside a widely used SDK's changelog is a signal worth tracking. Engineering teams that integrate with OpenAI's API through Vercel's AI SDK often gain early plumbing for models before they're publicly announced, since SDK maintainers coordinate with providers ahead of release. Developers should not attempt to call these model identifiers in production yet; treat their appearance as an indicator that a next-generation GPT-6 family may be approaching some form of availability, with official pricing, benchmarks, and documentation still to come from OpenAI directly.