Artificial intelligence news.
Straight from the original announcements.

Vercel AI Gateway Can Now Escalate Low-Confidence Answers to a Fallback Model

October 6, 2026By the AI News Bites Editorial TeamSource: Vercel
Vercel AI Gateway Can Now Escalate Low-Confidence Answers to a Fallback Model
Image: Vercel

Vercel's AI Gateway can now send a decision request to a fallback model when the primary model's answer trips a confidence condition set by the developer. The feature, called confidence-based decision fallbacks, is in beta.

Key points
  • Vercel's AI Gateway can escalate a decision request to a fallback model when the primary answer trips a confidence condition.
  • Conditions cover Choice and Score questions and can be combined; Boolean questions escalate on a probability range.
  • The fallback is set with one conditional object in providerOptions.gateway.models, and requests without it behave as before.
  • A triggered fallback runs a second decision and bills both stages.
  • Decision fallbacks are in beta.

How the new fallback works

Until now, fallbacks in AI Gateway have been about errors. According to Vercel, the plain model names already listed in models continue to catch outright errors, so existing fallback setups are unaffected. The new option adds a second trigger: an answer that comes back, but with too little confidence.

Developers define the condition, and the gateway escalates to a fallback model when the primary model's answer meets it. Vercel says conditions can be combined, so an escalation can depend on more than one signal.

Which question types are covered

Confidence conditions apply to Choice and Score questions. Boolean questions are handled differently: they escalate on a probability range instead of a confidence threshold.

The fallback is configured by adding one conditional object to providerOptions.gateway.models. In Vercel's example, the object names a model, openai/gpt-6-astra, and a when block with a question called intent and a confidenceBelow value of 0.6. In that setup, an answer to the intent question with confidence under 0.6 is escalated to the listed model.

Requests that do not include this object keep their existing behavior.

Scope of a condition and cost

The question field is optional. If it is left out, the condition checks every question of the matching type. Vercel gives the example { "confidenceBelow": 0.6 }, which escalates when any Choice or Score answer falls below 0.6.

There is a cost to consider. A triggered fallback runs a second decision, so it bills both stages, according to Vercel.

What it means for developers

The change lets teams use one model for decisions by default and bring in another only when the first is unsure, without changing how error fallbacks work. Because a triggered fallback is billed for both stages, the threshold a developer chooses affects how often the second charge applies. Vercel labels decision fallbacks as beta and points to its decision fallbacks documentation for details.

Related news

Alibaba’s Qwen4-Max arrives on Vercel’s AI Gateway
ModelsNovember 15, 2026

Alibaba’s Qwen4-Max arrives on Vercel’s AI Gateway

Alibaba’s Qwen4-Max is now available through Vercel’s AI Gateway, giving developers access to a reasoning model with native vision and a 1M-token context window. Vercel…

Products and tools
Products and toolsNovember 13, 2026, 09:00

Shield hands legacy IT portal tasks to OpenAI-powered agents

Shield says its Forge platform can now hand supported browser tasks to AI agents powered by OpenAI’s GPT-6 Astra, allowing IT providers to automate work in legacy…

Vercel Adds Kimi K3.1 Pro to Token Pass for Pro Teams
Products and toolsNovember 12, 2026

Vercel Adds Kimi K3.1 Pro to Token Pass for Pro Teams

Vercel has added Moonshot AI's Kimi K3.1 Pro to Token Pass, its Early Access subscription for Pro teams using coding agents through AI Gateway. The model supports…