Ampersand v0.1

humans& has released Ampersand v0.1, described as the first model trained for people to achieve more together, a collaborator that considers the downstream effects of its words and actions rather than only completing a task. The post argues mainstream training paradigms reward autonomy and task completion, which pushes models toward replacing work instead of enabling collaboration. The model is released through the humans& Agents API for developers building collaborative tools and apps. Evaluation embeds the model in a simulated group of users whose goals evolve over up to months of simulated time, scoring it on progress toward long-term goals minus the accumulated cost of interactions that conflicted with user preferences.

Vercel's November 6, 2026 changelog says Ampersand from Humans& is available on AI Gateway free through November 13, 2026, and applications call it as humansand/ampersand-v0.1 via the AI SDK, the OpenAI-compatible Chat Completions API, or a coding agent. The report is labeled Reported, meaning the authors did not run the model themselves. Access centers on two caller-executed tools: send\ message delivers a contribution to the conversation, while pass keeps the model quiet until the next human message. The design intent is that the model joins group conversations and decides whether to speak, offer clarifying questions, or share context. Pricing after November 13, 2026 and rate limits were not stated, and turn-taking quality was not measured.

humans& launched Ampersand v0.1 alongside Boba, a free assistant for one-on-one texting and group chats that went live on October 22, 2026 in the U.S., Canada and Australia. The model behind it is a post-train of Z.ai's open-weight GLM 5.3, which has 744 billion parameters, and it is trained to decide when to stay quiet rather than to maximize output. Developers can reach Ampersand v0.1 through an OpenAI-compatible Agents API with no waitlist, priced at $1.00 input, $0.20 cached input and $5.20 output per million tokens, with reasoning tokens included. All published benchmark figures for the model come from humans& and have not been independently verified, so they should be treated as vendor claims.

Ampersand v0.1 is a proprietary text model from humans& built for group collaboration, designed to work with people over long periods and to decide when to speak or stay quiet. BenchGecko reports it was released in October 2026 with a 1,048,576-token context window, text-in/text-out behavior, and is served through the humans& Agents API under the identifier ampersand-v0.1. The same listing says the model was post-trained from GLM 5.3 using reinforcement learning on long social simulations driven by the Persimmon user model, with roughly 744B total and 40B active parameters, and that weights are not released under a proprietary license. It also flags a notable gap: no benchmark scores have been recorded yet for the model.

A technical breakdown of Ampersand v0.1, released October 20, 2026, describes a model that joins a group conversation and after each message either sends one reply or passes, so it stays quiet most of the time. It is built on Z.ai's open-weight GLM 5.3 and is served through the humans& Agents API with no waitlist, following OpenAI's Agents API conventions so the official SDK works with their base URL. The analysis cites CooperBench, where GLM 5.3 alone scores 52.5% but drops to 17.3% when paired, while Ampersand reaches 39.0% with a partner, roughly 2.25 times the base model. It also reports that humans& priced the model at $1.00 per million input tokens and $5.20 per million output tokens, and that in humans&'s own simulation the model was the least likely of nine models to say it believed it was being evaluated.

Aggregator coverage confirms humans& published the full model card for Ampersand v0.1, a collaborator model trained to understand the long-term consequences of its actions and help groups pursue long-term goals together. It repeats the GLM 5.3 MoE architecture with 256 routed experts plus one shared expert, eight active per token, about 744B total and 40B active parameters, and a 1,048,576-token context. The report adds that Ampersand was trained with reinforcement learning on long-horizon open-ended social scenarios, with Persimmon simulating users in rollouts spanning hundreds of thousands of tokens while the model's actions reshape the conversation. It also notes dedicated sycophancy and evaluation-awareness evaluations, including a reported drop in eval awareness when Persimmon is used with test-time scaling.

The official Ampersand v0.1 model card lists the developer as humans& and the base model as GLM 5.3, a mixture-of-experts transformer with 78 layers, 256 routed experts plus one shared expert, eight active per token, roughly 744B total and 40B active parameters, and a 1,048,576-token context with text input and output. The version and API identifier are v0.1 and ampersand-v0.1, released October 20, 2026 via the Agents API. The model takes text conversation history, participant context, and optional tool results, and can emit messages, tool requests, or stay silent. It was post-trained with reinforcement learning on long-horizon, open-ended social scenarios, with the Persimmon user model playing simulated users across rollouts spanning hundreds of thousands of tokens. The card also flags sycophancy and evaluation-awareness tests, and states applications must enforce permissions and validate tool actions.