humans& releases Ampersand v0.1, a 'collaborator' model RL-trained from GLM 5.3 to help groups reach shared goals, via an Agents API

Why it matters

【Key points】

humans&, a lab founded by alumni of OpenAI, Anthropic, Google DeepMind and xAI ($480M seed in January), released its first model, Ampersand v0.1: a "collaborator" that sits in group chats and team tools and helps people reach individual and shared goals instead of doing tasks on its own.

1. Base: RL on Z.ai's GLM 5.3 (MoE, ~744B total / ~40B active, 1M context), text only. Simulated users came from humans&' own user model, Persimmon, over months of simulated time, with a new off-policy RL method (CORO).

2. Self-reported, tentative results: long-term collaboration sim 12.94 vs 9.89 for GLM 5.3; Collaborative Intelligence 59.1% (GLM 5.3: 54.3%); CooperBench cooperative coding 39.0%, vs GLM 5.3's 17.3%. Trade-offs: higher interaction cost (it coaches when people want answers), and an SWE-bench Pro subset dips from 84.3% to 82.4%.

3. Safety: WMDP-cyber rises from 80.6% to 89.0%; prompt-injection rates stay near GLM 5.3, which humans& says is well above frontier models; refusals and sycophancy improve.

4. Access: Agents API now, no waitlist, OpenAI Agents API-compatible. $1 input / $0.20 cached / $5.20 output per million tokens. No open weights.

【Verifiable】

Blog: https://ampersand.humansand.ai/blog/

Model card: https://ampersand.humansand.ai/blog/model-card.html

Pricing: https://ampersand.humansand.ai/docs/PRICING.html

Official X post: https://x.com/humansand/status/2112674550907552137

【Our read】

1. The new part is the objective: it is scored on how far each person's goals moved over long group interactions, minus unwanted interruptions, and it is trained to stay silent (send or pass after every message). That cuts against the push toward ever more autonomous agents.

2. All numbers are self-reported, and the headline eval uses humans&' own design and its own simulated users.

3. It is a GLM 5.3 fine-tune with more cyber hazard knowledge and weak injection defenses; group-chat integrations must enforce permissions and check tool calls.