Humans& reports 39.0% on CooperBench. The benchmark's public page lists GPT-5 and Claude Sonnet 4.5 at 25% and shows no Ampersand result. The company also says the model's interaction costs are higher than its peers'. Our hypothesis is that the release is about staking out a metric. The simpler explanation is that it is about collecting real usage data early. Either way, the numbers can't be checked from outside yet.
Summary
Humans& has released Ampersand v0.1 through its Agents API. It calls the model a "collaborator" trained with reinforcement learning in long-horizon, multi-user simulations, not on verifiable or preference rewards. The company reports 59.1% on Collaborative Intelligence at high effort and 39.0% cooperative success on CooperBench. It also says Ampersand's interaction costs, including unwanted coaching, are higher than those of other leading models. Humans& is the source of every number, and nobody has replicated them independently.
What happened
The release post says Ampersand was trained inside simulated groups with seeded goals that can change over episodes lasting up to months of simulated time. According to the model card (d1), rollouts run to hundreds of thousands of tokens. A separate Humans& model called Persimmon plays the users. Humans& describes Persimmon as a simulator of people in multi-turn, multi-user chat whose users can push back, resist sycophantic behavior and change their goals (d2). The objective scores progress toward users' long-term goals minus interaction costs that conflict with their preferences. The post presents this as a correction to post-training that optimizes task completion or shallow preferences, which it says pushes models toward replacing work instead of enabling collaboration. The headline results are 59.1% on Collaborative Intelligence (high effort) and 39.0% cooperative success on CooperBench. In the same release, Humans& says Ampersand's interaction costs are higher than those of other leading models and names unwanted coaching as one source. Developers building collaboration tools can use the model through the API now. The release includes no pricing, latency or deployment results.
What the archive adds
The archive shows two neighboring patterns. OpenAI's case study on Amp (Oct 14) describes a single-user agent that grows beyond code, running on OpenAI models including GPT-6 Astra. That is the autonomous-worker pattern Humans& says it is moving away from. Odyssey's Agora-2 post treats multi-agent simulation as training ground, and it notes that failures found during agent interactions become training data for its world model. Odyssey simulates physical and game worlds, though, not people. It is a parallel to the Humans& approach, not the same bet. On the benchmarks, according to CooperBench's public page, GPT-5 and Claude Sonnet 4.5 sit at 25% and no Ampersand result is listed. Collaborative Intelligence evaluates on a private held-out set. Readers therefore can't currently see how the 59.1% was produced or verified. The archive also holds a LangChain post about Included Health's Dot, a multi-agent healthcare guide that, per that post, uses multi-turn simulation to evaluate the agent (d3). This suggests simulated users are becoming a normal evaluation tool, not something only Humans& does.
Trend signal
Two patterns meet here. The first is labs defining the axis they compete on. Frontier models are compared mostly on reasoning, coding and autonomy. Humans& proposes a different scoring rule, group goal progress minus intervention cost over long horizons, and publishes against it first. The second is simulated users moving from evaluation into training. Dot uses multi-turn simulation to evaluate (d3), Odyssey trains on simulated worlds, and Humans& trains directly against a people simulator (d1, d2). When rewards are verifiable, the grader is a test suite. In social RL, the grader is a model of people, and its blind spots become the policy's blind spots. We can't yet say whether people simulators will be scarce or easy to commoditize. Our evidence is a handful of releases, not a market.
Why now
1) A stated objection to autonomy-focused post-training. The release argues that task-completion and preference rewards push models to replace work, and the model card frames long-horizon social RL as the alternative (primary; d1). 2) User simulators built to resist compliance. Persimmon is described as pushing back, resisting sycophancy and changing goals, which makes RL against simulated people plausible (d2). 3) Teamwork benchmarks that a small lab can publish against: CooperBench and Collaborative Intelligence, the latter with a private held-out set (primary). 4) Simulation is spreading as an agent evaluation method in production settings such as Included Health's Dot (d3), which lowers the cost of accepting simulation-based claims.
Connections to other events
Why ship a v0.1 the company itself says coaches too much? We see two readings and can't settle between them.
Hypothesis A, metric-setting: the release stakes out "progress minus interaction cost" as the way team agents should be graded, before any larger lab picks a framing. The comparison set fits this reading. The 39.0% on CooperBench looks like a 14-point lead only if the baselines are the GPT-5 and Sonnet 4.5 figures on the public page. That page lists no Ampersand entry, and the release doesn't make clear which model versions it compared against. On Collaborative Intelligence, the held-out set is private, so outsiders can't reproduce the 59.1%. Publishing the interaction-cost weakness also fits this reading: a metric that subtracts cost looks more honest when its author reports its own cost.
Hypothesis B, data collection: a research lab ships early because simulated users only go so far, and API traffic from real groups is the only way to calibrate Persimmon. This is the simpler explanation, and it predicts fast iteration over benchmark promotion.
The unwanted coaching may also follow from the objective itself. This is our inference, not a Humans& claim. If the reward is goal progress and the simulated users can be steered, then intervening (nudging, suggesting, coaching) is often the fastest way to move the simulated group forward. The interaction-cost term pushes back only as far as Persimmon's users register annoyance. If the simulator under-penalizes unsolicited advice compared with real people, the policy learns to over-coach. A field study of a proactive "Team Agent" in a workplace, cited in our research notes, describes this kind of friction. That suggests the gap between simulated and real tolerance is where this design will be tested.
Why it matters
Builders of chat, project-management or group tools have a model whose objective matches what they care about, but its admitted weakness is the behavior their users dislike most. Count unsolicited interventions as a cost you budget for, not as proof of collaboration. Buyers shouldn't read 59.1% or 39.0% against capability leaderboards. The numbers say nothing about coding or reasoning strength, so a likely pattern is routing: a frontier model does the task work and a collaboration-tuned layer handles coordination. If Hypothesis A is right, the next signal is other vendors starting to report multi-user metrics. If Hypothesis B is right, the next signal is a quick v0.2 that cuts interaction cost, with little benchmark marketing.
What happens next
- A party other than Humans& publishes Ampersand results on CooperBench or Collaborative Intelligence within 5 points of the self-reported 39.0% / 59.1%. (6 months, 35%). Watch for: An Ampersand entry appears on the public CooperBench page or in a third-party eval report.
- Humans& ships a successor model whose release reports lower interaction cost or less unwanted coaching than v0.1 on a named metric. (6 months, 55%). Watch for: A new model card with an interaction-cost or unsolicited-intervention comparison against v0.1.
- OpenAI, Anthropic or Google includes a multi-user collaboration benchmark (CooperBench, Collaborative Intelligence or similar) in an official model release. (12 months, 30%). Watch for: A frontier launch post or model card with a cooperative or multi-user results table.
Limits and uncertainty
Humans& self-reports all Ampersand figures. We don't know the exact baselines or versions behind either comparison, so the 14-point reading depends on the CooperBench public page and could be wrong if Humans& tested current frontier models. The Collaborative Intelligence held-out set is private, so we can't comment on how robust the score is. Persimmon's fidelity to real people is asserted, not shown. The over-coaching mechanism above is our inference from how the objective is described, not a finding from Humans&. The strategic reading may be overdone: Hypothesis B, early data collection, fits the evidence at least as well. We have no pricing, latency or deployment outcomes.
Skill to build now
Build a cost-penalized eval for multi-user agents and stress-test the cost term. First exercise: take one real team thread of about 50 messages and script three simulated participants with distinct goals, one of whom changes their goal halfway. Run two variants of the simulator, one tolerant of advice and one that gets irritated by unsolicited messages. Log goal progress and unprompted agent messages for each. If the agent's score ranking flips between the variants, your intervention penalty is too dependent on how the simulator is built.
Reader takeaway
If you build a multi-user product, run your own small bake-off before you trust any collaboration score. Replay 20 anonymized real group threads through Ampersand and through your current model. For each, log whether the group's actual decision or deliverable advanced and how many unsolicited agent messages the humans ignored or rejected. Use Ampersand for coordination only if it wins on progress without a higher rejected-intervention rate.
Open question
Is a model trained against Persimmon learning how real groups collaborate, or learning how much intervention Persimmon tolerates? Can anyone outside Humans& tell the difference without access to the simulator or the private held-out set?