Ampersand on Vercel: free window and direct API costs
Vercel’s Ampersand promotion has a date, not a promise of permanent free access. Its November 6 announcement says the group-conversation model is free through November 13, 2026. The current AI Gateway listing still shows free input and output. Neither page establishes an exact cutoff time or a post-promotion price.
That matters for developers comparing two different routes: Vercel’s gateway and the paid Agents API operated by humans&. Ampersand itself launched on October 20. This is an access and integration guide, not a new model announcement.
What Ampersand does differently
Ampersand is intended to participate in conversations involving several people, including deciding when not to contribute. In Vercel’s integration example, the application supplies send_message and pass function tools. The application, not the model alone, executes those actions and records their results.
A useful evaluation is therefore not just whether its replies sound helpful. Check whether it identifies the right recipient, remembers a relevant constraint, and leaves a conversation alone when no contribution is needed. Those are suggested tests, not results from ByteForward testing.
Vercel AI Gateway access
Use humansand/ampersand-v0.1 through the AI SDK or OpenAI-compatible Chat Completions route. Vercel also documents setup for coding agents, but that availability is not evidence that a group-conversation model outperforms a coding specialist.
The current model listing specifies a shared one-million-token context for prompt and response, with an output ceiling of 32,768 tokens including reasoning. Check that listing and your account’s terms before relying on the temporary free offer for a production budget.
The direct API has separate prices
The humans& pricing page lists US dollar rates per million tokens: $1 for uncached input, $0.20 for cached input and $5.20 for output, including reasoning. These are direct API rates, not a verified forecast of Vercel’s next price.
Cache hits are not guaranteed. Tool-result continuations and context compaction can add model usage, while the application pays for its own custom tools separately. The Normal tier lists 30 API requests and 10 work requests per minute, with 16 concurrent executions. Higher account limits do not change the published token rates.
Persistent sessions are not a turnkey group chat
In the direct API guide, a session stores conversation history and a copy of agent settings. Applications send later messages and tool results to that session, then read streamed events or saved state. They still decide where replies appear and implement the functions the agent requests.
The optional multi-agent mode is different from several human participants: it lets one session use subagents of the same agent. The guide recommends leaving it off for ordinary chat and short tasks, and says every subagent adds usage. Do not assume a gateway call exposes every direct-session feature.
What to check before adding it to a team
The October 20 model card describes a text model based on GLM 5.3 and cautions that simulation results may not generalize to real groups. It can misunderstand goals, interrupt or disclose inappropriate context.
Keep participation visible, obtain appropriate consent, enforce access controls and require human confirmation for consequential actions. A model choosing to speak is not permission to reveal a private message. For a pilot, measure unwanted interruptions and recipient mistakes alongside useful contributions, and budget separately for the route you actually use.
For a consumer-facing alternative rather than an API integration, our Google Tabletop access guide covers a shared planning workspace with a waitlist. The distinction is important: joining a collaboration service is different from building one around a model.
Featured image: AI-generated editorial illustration, not a product screenshot. Vercel, the Vercel design, Next.js and related marks, designs and logos are trademarks or registered trademarks of Vercel, Inc. or its affiliates in the US and other countries.
