Ampersand Launches on AI Gateway: The Group Chat Language Model That Knows When to Stay Silent

AI assistants in group chats usually share one common flaw: they talk too much. Whether or not they're needed, they have to chime in. Humans&'s Ampersand tries to solve this problem—it's been trained to actively judge "should I speak or should I shut up" at every turn of the conversation. The model is now available through Vercel AI Gateway and free to use until November 13, 2026.

A Model Built for Group Chats

Ampersand isn't positioned as a general-purpose conversational assistant, but as a participant in multi-person group chat scenarios. It needs to understand the difference between individual goals and shared team goals, help move discussions forward when people disagree, and judge whether its own intervention is worthwhile.

That intervention might be asking a clarifying question, adding relevant context, or simply waiting for others to respond first. In other words, "not speaking" is designed as a formal output option, not a failure state for the model.

According to the official description, Ampersand's core capability lies in recognizing "whether a contribution is useful right now." This judgment is harder to evaluate than mere response quality, and it's also closer to how humans actually behave in real collaborative scenarios.

Two Function Tools: send_message and pass

Ampersand's behavior is exposed through two function tools defined on the application side. send_message delivers a group chat message, while pass means staying quiet until the next human message arrives. The execution and result logging of both tools are handled by the calling application.

Tool selection needs to keep the default auto setting. This is worth noting: Ampersand accepts auto and none, but explicitly rejects required and named tool selection. This design limits developers' ability to force the model to call a specific tool, consistent with the product philosophy of "letting the model decide whether to speak."

For integration, you can use the AI SDK, the OpenAI-compatible Chat Completions API, or a coding agent. The model name is uniformly humansand/ampersand-v0.1.

Function tools are a tool-calling mechanism introduced by OpenAI: developers declare a set of callable functions using JSON Schema, the model decides during inference whether to call one, which one, and with what parameters, and then the application actually executes it and passes the result back to the model. Unlike traditional structured output, tool calling separates "model decision" from "actual side effects," making it especially suitable for scenarios that require external actions. Ampersand leverages this mechanism to model both "speaking" and "staying silent" as tools, so that "not calling a tool" is no longer an implicit default behavior but an action that can be explicitly logged and validated.

toolChoice is a parameter that controls the model's tool-calling strategy. Common values include auto (the model decides whether to call tools), none (prohibits tool calls), required (must call at least one tool), and specifying a particular tool name (forces that tool to be called). Ampersand only accepts the first two and rejects the latter two, effectively prohibiting the caller from "commanding" the model to speak at the protocol level—consistent with its design of leaving the decision to speak to the model itself.

Key Constraints in the Integration Code

The official example shows a complete AI SDK integration flow: maintain a message history array, define two tools, and trigger a streamText call whenever a human message arrives. The example sets maxRetries: 0, toolChoice: 'auto', stopWhen: stepCountIs(1), and disables parallel tool calls.

The validation logic in the code is quite strict: the finish reason must be tool-calls, the number of tool calls must be exactly 1, and it must not be a dynamic tool. Any case that doesn't match expectations throws an error. This reflects Ampersand's design assumption of "one decision per turn"—either speak or stay silent; there's no semantics for executing multiple actions simultaneously.

The two-turn conversation in the example illustrates the point well. The first human message invites the model to participate ("Ask one clarifying question"), while the second explicitly asks it to wait for Leo to reply. pass is handled by logging the wait action—neither publishing a message nor triggering the next model call.

On the engineering side, the official guidance recommends persisting responses, history, and delivery receipts to reliable storage; retrying uncertain deliveries using the saved action and the same receipt; processing messages in session order; and checking session access permissions before publishing. These recommendations point to a real-world problem: concurrency and retries in group chat scenarios are far more complex than single-turn Q&A.

The stopWhen: stepCountIs(1) in the example is part of the AI SDK's multi-step generation control. By default, after the model returns a tool call, the SDK executes the tool and feeds the result back to the model, triggering the next generation step and forming a loop of "tool call—result feedback—re-inference." stepCountIs(1) limits the loop to one step, meaning it stops as soon as the tool call is received and no longer automatically initiates the next model request. This maps exactly to Ampersand's "one decision per turn" semantics and explains why pass doesn't immediately trigger a new model call—the continuation of silence is driven by the next human message, not by the model's self-looping.

What AI Gateway Provides

AI Gateway itself is a unified calling layer that tracks usage and costs, configures retries and failover, and provides performance optimizations that exceed the availability of any single provider. It includes capabilities such as custom reports, API key budgets, and routing rules.

On pricing, AI Gateway bills at provider prices with no markup, charges no platform fee for inference, and the same applies to BYOK requests. For teams wanting to experiment with Ampersand at low cost, the free window plus the zero-markup policy lowers the cost of trial and error.

When using it in a coding agent, you first need to install the latest version of the Vercel CLI and run npx vercel ai-gateway setup, then select humansand/ampersand-v0.1 in the agent. The model playground also provides a direct entry point for trying it out.

BYOK (Bring Your Own Key) means developers use their own API keys obtained from model providers and route requests through the gateway. In this mode, the gateway typically only handles routing, observability, and failover without touching the provider's billing; AI Gateway states that BYOK requests are also not marked up, meaning teams can continue to use their existing contracts and quotas while gaining unified usage tracking and routing capabilities. For teams that already have accounts with multiple providers and want centralized key and cost management, this is a smoother path when integrating new models like Ampersand.

Opportunities and Concerns for This Kind of Model

Making "silence" a first-class citizen is a reasonable direction of evolution for group chat AI. Existing assistants often create noise in multi-person conversations, while Ampersand leaves the decision to speak to the model itself, which in theory can make collaboration more natural. It has direct appeal for enterprise chat, project discussions, multi-person decision-making, and similar scenarios.

However, "judging when to speak" is inherently subjective and context-dependent. Whether the model's understanding of shared goals is accurate, and whether its interventions in disagreements are truly helpful, can only be verified through actual deployment. The official example uses a system prompt requiring it to "choose one action," which shows that model behavior still depends heavily on the caller's prompt design.

Additionally, tool selection rejecting required and named modes, while consistent with the product philosophy, also means developers cannot force the model to output a message in specific scenarios. For applications that require deterministic behavior, this is a constraint that needs to be evaluated in advance.

Ampersand is currently in the v0.1 stage, and the free period gives developers a window to observe. Whether it can become a reliable participant in group chat scenarios depends on whether its silence judgments in real conversations are restrained enough and accurate enough.

One-line Summary

Humans& launches Ampersand, a model designed specifically for group chats that makes "silence" a formal output option.

Paragraph Summary

Ampersand is a group-chat-specific language model from Humans& whose core innovation is designing "not speaking" as a formal output just as important as speaking. It exposes its behavior through two function tools, send_message and pass, actively judging each turn whether intervention is worthwhile. It requires tool selection to remain on auto and rejects required and named tools, preventing developers from forcing the model to speak at the protocol level. The model is named humansand/ampersand-v0.1 and can be integrated via the AI SDK, the OpenAI-compatible Chat Completions API, or a coding agent. The official example strictly validates the finish reason and the number of tool calls, reflecting the "one decision per turn" design assumption. The model is now available through Vercel AI Gateway, billed at provider prices with no markup, and free until November 13, 2026. Its opportunity lies in reducing group chat noise, while its concern lies in the subjectivity of silence judgments and its dependence on prompt design.