Ampersand by Humans& official social card for the collaborator language model

Ampersand v0.1: Power-User Guide to the Collaborator Model

by

Ampersand v0.1 from Humans& is a collaborator language model for group chats that chooses send_message or pass each turn. Use the Humans& Agents API directly, or call Vercel AI Gateway as humansand/ampersand-v0.1, which remains free through November 13, 2026 for power users wiring team chat agents.

Last updated: November 7, 2026

Table of contents

What Ampersand v0.1 is for

Humans& describes Ampersand as the first model trained for people to achieve more together. Instead of always completing a task alone, it is meant to join a team chat, track individual and shared goals, help with disagreements, and recognize when a clarifying question, shared context, or silence is the useful move. The company published the model through its Agents API (blog dated October 20) and Vercel made humansand/ampersand-v0.1 available on AI Gateway on November 6, 2026.

Training started from Z.ai’s GLM 5.3. Humans& reports stronger Collaborative Intelligence and CooperBench results versus that base, plus multi-agent coding scores on SWE-bench Pro v2 HARD-51 (92.2% in their multi-agent Agent API setup) and DeepSWE v1.1 (77.9%). Treat those as vendor-reported figures from the Ampersand announcement. For other agent deploy patterns, see Mistral Managed Agents and Hermes Agent 0.23.

Item Official detail Source
Model IDs ampersand-v0.1 (Humans&); humansand/ampersand-v0.1 (AI Gateway); tutor ampersand-tutor-v0.1 Docs / Vercel
Role Group collaborator; speak or stay quiet each turn Blog / changelog
Context / max out 1,000,000 context; 32,768 max output (gateway page) Vercel model page
Core tools Caller-defined send_message and pass Docs / changelog
Gateway promo Free on Vercel AI Gateway through November 13, 2026 Vercel changelog
Token prices (Humans&) $1.00 / $0.20 cached in; $5.20 out per 1M Pricing page

Access paths: Humans& API and Vercel AI Gateway

On Humans&, sign in at ampersand.humansand.ai for an API key with no waitlist. Point the official OpenAI Python client at https://ampersand.humansand.ai/v1, set model ampersand-v0.1, and use the Agents API (OpenAI-Beta: agents=v1). Create one session per conversation so history lives server-side; later turns only need new messages and tool results. Docs recommend openai==3.24.0 and max_retries=0 on create so a lost response is not retried into a second session.

On Vercel AI Gateway, call humansand/ampersand-v0.1 with the AI SDK or Chat Completions against https://ai-gateway.vercel.sh/v1. The November 6 changelog states free access through November 13, 2026. Coding agents (Claude Code, Codex, Cursor, and similar) can select the model after npx vercel ai-gateway setup. Related FreeAnswers reading: Open Duet 0.8 for self-hosted coworkers and Claude Message Threads for API conversation continuity.

send_message, pass, and chat integration

Ampersand was trained so chat actions are function tools, not markup in free text. Declare send_message (posts a contribution) and pass (yields without a visible message). Your application checks channel access, publishes or skips, then returns the tool result with the same turn_id and call_id. Do not also auto-publish ordinary assistant text after a tool already posted.

Gateway samples set tool_choice to auto, disable parallel tool calls, and expect exactly one completed tool action per human message. Ampersand accepts auto and none but rejects required and named tool choices. Persist responses, history, and delivery receipts; retry uncertain delivery with the same receipt; process messages sequentially per conversation. For personal versus group routing patterns, Humans& documents separate sessions per assistant and application-enforced membership—similar in spirit to multi-surface agents covered in OpenClaw 2026.10.10 and Muse Spaces.

Optional agent.multi_agent on ampersand-v0.1 runs a main agent plus up to N subagents of itself (default 3) for long coding-style work. Wait for agent.session.idle. Reasoning effort defaults to high (16,384-token output budget); low, medium, and max (32,768) are documented for the standard model.

Pricing, tiers, and limits

According to the Humans& pricing page, both public models use the same token rates: $1.00 per million uncached input tokens, $0.20 cached input, and $5.20 output including reasoning. Usage covers continuations after tool results and context compaction. Normal accounts allow 30 API requests/minute, 10 work requests/minute, 16 concurrent executions, and 10,000 sessions; High raises those to 600 / 300 / 64 / 1,000,000. Gateway promotional free access does not replace Humans& account pricing after the promo window.

Hosted input caps reserve space for the chosen effort (for example 241,664 full-model input tokens at high effort). Compaction can be auto or disabled. New ampersand-v0.1 agents default allow_silent_completion to true; pair a pass tool with allow_silent_completion=false when every turn must end in an explicit speak-or-pass choice.

How to wire Ampersand into a group chat

  1. Create an API key at ampersand.humansand.ai, or an AI Gateway key for the free window, and keep the key on your backend only.
  2. Declare send_message and pass function tools; keep tool choice at auto (or none). Do not ask the model for XML-style reply markup.
  3. Open one Agents API session per conversation (model ampersand-v0.1) or call humansand/ampersand-v0.1 on the gateway with a durable message list.
  4. Send every human message with sender metadata (sender, channel, optional timestamp) so Ampersand knows who said what.
  5. On each human message, run one model turn, execute the single tool call, save a delivery receipt, then append the tool result before the next message.
  6. For production, process conversations sequentially, verify channel access before posting, and reconnect from saved session/turn state if SSE drops.

Frequently asked questions

What is Ampersand v0.1?

Ampersand v0.1 is Humans&’s collaborator model for group conversations. On each turn it decides whether to speak (send_message) or stay quiet (pass), with training focused on shared goals, timing, and restraint rather than always answering.

How do I call Ampersand on Vercel AI Gateway?

Use model ID humansand/ampersand-v0.1 through the AI SDK, the OpenAI-compatible Chat Completions API at https://ai-gateway.vercel.sh/v1, or a coding agent after npx vercel ai-gateway setup. Vercel lists Ampersand free on AI Gateway through November 13, 2026.

What do the send_message and pass tools do?

Your app declares both as function tools. send_message posts a contribution; pass finishes the turn without a visible message. Your app executes the chosen tool, records the result, and should leave tool_choice at auto or none (required and named choices are rejected on the gateway path).

How much does Ampersand cost on the Humans& API?

Official pricing is $1.00 per 1M uncached input tokens, $0.20 per 1M cached input, and $5.20 per 1M output tokens (including reasoning). Normal and High tiers share those rates and differ on rate, concurrency, and session limits.

What is the context window for Ampersand v0.1?

Vercel’s model page lists a 1,000,000-token context window and up to 32,768 maximum output tokens for humansand/ampersand-v0.1. Humans& docs also publish per-effort full-input token caps (for example 241,664 tokens at high effort) that reserve room for the output budget.

Can Ampersand run multi-agent coding in one session?

Yes on ampersand-v0.1. Set agent.multi_agent.enabled to true and optionally max_concurrent_subagents (default 3). Subagents share the session’s tools and effort; wait for agent.session.idle when every agent is done. The tutor model rejects multi-agent mode.

Sources

Related answers

Leave a Comment