Vercel AI SDK shipped `@ai-sdk/[email protected]` on 2026-09-26: request-level and positioned system-message `reasoningEffortUpdate` now accept `'none'` so GPT-6 Sol/Luna conversations can turn reasoning off mid-thread. Updates are validated against each model’s supported efforts—unsupported request-level updates warn and omit; unsupported historical/message-level updates reject. Companion `@ai-sdk/[email protected]` and `@ai-sdk/[email protected]` pick up the dependency.
Key Takeaways
- ✓Shipped: @ai-sdk/[email protected] / npm 4.0.78
- ✓Feature: reasoningEffortUpdate 'none' for GPT-6 Sol/Luna (#21518)
- ✓Validation: warn+omit unsupported request-level; reject unsupported historical updates
- ✓Companions: Azure 4.0.82, Bedrock 5.0.97; see OpenAI provider docs
- ✓Install: npm i @ai-sdk/[email protected]
Project Links & Resources
Direct AccessIn-Depth Technical Analysis
Core Background & Industry Pain Points
GPT-6 Sol/Luna can set reasoning effort to none mid-conversation while preserving the prompt-cache prefix. The SDK previously rejected 'none' in reasoningEffortUpdate, so apps could not turn reasoning off after earlier turns used it.
Architecture Highlights & Internals
@ai-sdk/[email protected] (94d5d6d) allows 'none' for request-level and positioned system-message updates, validates against each model’s supported efforts (warn+omit vs reject), and ships with Azure 4.0.82 / Bedrock 5.0.97 dependency bumps. Builds on message-level updates from #21418.
Authoritative Benchmarks & Measured Scores
No latency/cost benchmarks published; PR notes live provider acceptance was not run. Measure reasoning-token and bill deltas for low→none on a fixed multi-turn script while confirming cache prefixes stay intact.
Developer Hands-on Guide
npm i @ai-sdk/[email protected]; set providerOptions.openai.reasoningEffortUpdate: 'none' per OpenAI provider options. npm: 4.0.78.
Discussion & Comments
0Sign in to join the discussion
Connect with AI developers to exchange benchmark insights.