Humans&'s Ampersand model for group-chat turn-taking is free on Vercel's AI Gateway

Vercel has added Ampersand, a language model from Humans& built to take part in group conversations, to its AI Gateway. According to the Vercel blog, the model is available free of charge through November 13, 2026.

A model that decides when to speak

Unlike a typical chat model that responds to every incoming message, Ampersand is designed to sit inside multi-participant settings such as a team chat and decide, turn by turn, whether to contribute or stay quiet. Vercel says the model is built to understand individual and shared goals among participants, help people work through disagreements, and recognise when a contribution would actually be useful. That contribution might be a clarifying question, a piece of relevant context, or nothing at all while people respond to one another.

How the turn-taking works

Ampersand exposes its decision as caller-defined function tools rather than keeping it internal. Two actions are defined: send_message delivers a contribution, and pass leaves the conversation quiet until the next human message arrives. The calling application is responsible for executing both tools and recording their results.

Vercel notes a configuration constraint: Ampersand accepts the auto and none tool-choice settings but rejects required and named tool choices, so developers should leave tool choice at its default.

The model can be called through the AI SDK, the OpenAI-compatible Chat Completions API, or a coding agent, using the model name humansand/ampersand-v0.1. The intended pattern is message-driven: the first human message invites a contribution, while the second asks Ampersand to wait. In a chat application, Vercel says to call the handler whenever a new human message arrives. When the model returns pass, the application records the waiting action without posting a message or triggering another model turn.

Production guidance

For deployed chat applications, Vercel recommends persisting the response, the conversation history, and delivery receipts in durable storage. If delivery is uncertain, the retry should reuse the saved action and the same receipt. Vercel also advises processing messages sequentially per conversation and checking conversation access before posting anything.

Ampersand can also be used from coding agents such as Claude Code, Codex and Cursor. After installing the latest Vercel CLI and running npx vercel ai-gateway setup, developers select humansand/ampersand-v0.1 in the agent. There is also a model playground for trying it out directly.

Access through AI Gateway

AI Gateway is Vercel's unified API for calling models, tracking usage and cost, and configuring retries, failover, and performance optimisations aimed at delivering higher uptime than individual providers. It also includes custom reporting, budgets for API keys, and routing rules. Vercel says the gateway reflects provider pricing with no markup and charges no platform fee on inference, including on Bring Your Own Key requests.

Why it matters

Most LLM integrations assume a one-to-one conversation in which the model always replies. Group settings such as team chats, shared threads, and meeting backchannels break that assumption, and a model that answers every message quickly becomes noise. Ampersand treats speaking as an explicit decision rather than a default, an approach that could make AI participants viable in the multi-party spaces where much real work happens. Because it is exposed through a standard OpenAI-compatible endpoint and costs nothing during the trial window, developers can test turn-taking behaviour with minimal friction. The v0.1 version string signals an early release, and the free period runs only through November 13, 2026, so teams evaluating it should note that pricing and behaviour may change after that date.