humans& Says Its Ampersand Model Beats Claude Fable for Groups, but Its Own Chart Data Puts Them 0.002 Points Apart and GPT-6 Astra Leads on Teamwork

humans&, the AI lab that came out of stealth in January with backing from NVIDIA, Jeff Bezos and GV, released its first model on 20 October 2026. Ampersand v0.1 is pitched as a “collaborator” that sits in group chats and workspaces and decides, message by message, whether to speak or stay quiet.

The launch post on X says users working with Ampersand “accomplished more than with Fable” in multi-month simulations. It had passed 80,000 views by Wednesday morning IST. Access is open now through an API, with no waitlist, at $1.00 per million input tokens and $5.20 per million output tokens.

We read the release blog, the model card, the chart data files on the release page and the public leaderboard of the Stanford benchmark humans& cites. The Fable claim holds on raw goal progress. Once humans& applies its own penalty for annoying users, the two models are 0.002 points apart.

What Ampersand actually is

Ampersand is not a new base model. The model card says it is built on GLM 5.3, the open-weight model from Z.ai, and keeps that architecture: a mixture-of-experts design with about 744 billion total parameters, about 40 billion active per token, and a 1,048,576-token context window. It takes text in and sends text out.

What humans& changed is the training. It ran reinforcement learning inside long simulations where Persimmon, a user model the company previewed in September, plays the people. Those simulated users have hidden goals, push back on unhelpful replies and change their minds over time. The blog also describes a new off-policy method it calls CORO and an NVFP4 training recipe run on B300 GPUs.

The product idea is simple. Developers give Ampersand two tools, send_message and pass, and after every group message it picks one. In the company’s demo, it stays silent through two messages, then flags that a Friday rehearsal will not work because a teammate said earlier he would be away.

Pricing, limits and how it compares with GLM 5.3

The API follows OpenAI’s Agents API, so the official OpenAI SDK works by pointing it at the Ampersand base URL. Billing is prepaid credit. Reasoning tokens count as output even though the reasoning text is not returned.

Per 1M tokens (USD)	Ampersand v0.1	GLM-5.3 on Z.ai
Input	$1.00	$1.40
Cached input	$0.20	$0.26
Output (incl. reasoning)	$5.20	$4.40

So Ampersand is cheaper to read with and dearer to write with than its own base model on Z.ai’s price list. For a group chat bot that reads a lot and posts little, that mix favours Ampersand. For long reasoning-heavy replies, it does not.

The Normal tier allows 30 API requests a minute and 16 concurrent executions. One limit is easy to miss: Normal accounts can create 10,000 sessions in total, the count never resets, and deleting a session does not free a slot. Since the docs recommend one session per conversation, a busy consumer app could hit that ceiling quickly. The High tier raises it to 1,000,000.

The Fable claim, checked against humans&’ own chart data

The headline test, Collaboration Over Time, drops each model into a simulated social group chat for up to 77 simulated days. The score is goal progress minus a weighted “interaction cost” for things users did not want, like unsolicited coaching. The release page publishes the underlying numbers in a data file, along with a slider that lets readers change the cost weight. Its default for this chart is 0.2.

Model	Goal progress	Interaction cost	Score at default weight
Claude Fable 5.1	13.45	2.50	12.945
Ampersand v0.1	13.82	4.37	12.943
Gemini 3.8 Flash	13.29	1.85	12.924
Claude Opus 5.5	10.83	1.33	10.56
GLM 5.3 (base)	10.97	5.40	9.89
GPT-6 Astra	9.96	0.76	9.80
GPT-6.1 Sol	9.20	0.98	9.00

Ampersand does make the most goal progress, which is what the X post says. But its interaction cost is the second highest of the nine models, behind only GLM 5.3, and about 5.7 times GPT-6 Astra’s. At the default weight, Fable, Ampersand and Gemini 3.8 Flash land within 0.03 points of each other, while the standard errors on each are around 0.8 to 0.95. Double the weight to 0.4 and Ampersand drops to fourth.

The blog is open about the cause. “Ampersand sometimes offers guidance or encouragement when people simply want an answer or prefer to work things out themselves,” it says. It also notes GPT-6 Astra “rarely does anything against users’ preferences.” Two smaller caveats: notes in the data file show GPT-6 Astra, GPT-6.1 Sol and Claude Opus 5.5 were scored on 71 or 72 cases, against 108 for Ampersand, and the model card labels all results “tentative” and says they “may change as runs finish.”

The fair reading is that a GLM 5.3 fine-tune now matches the best closed models on this social test, up from about 9.9 for the base model. That is a real jump. It is not a clean win.

Teamwork: best open-weight result, still behind the closed leaders

For teamwork, humans& used the Collaborative Intelligence benchmark from James Zou’s group at Stanford. It scores how often a model still passes a task when a teammate causes a problem. humans& reports 59.1% at high effort and compares it with “the strongest open-weight public-board result” of 49.8%, which belongs to DeepSeek V4.1 Flash.

That comparison is accurate but narrow. On the benchmark’s public board, last generated 16 October, closed models sit well above Ampersand: GPT-6 Astra at 77.7%, GPT-6.1 Sol at 74.2%, Claude Opus 5.5 at 70.4% and Claude Fable 5.1 at 66.6% on the released scenarios. Ampersand is not on that board. Its score comes from humans&’ own run of the 160 released scenarios, at high effort, while the board models were run once at their lightest reasoning setting, according to humans&’ own notes.

The per-dimension data shows trade-offs the blog skips. Against GLM 5.3 run the same way at high effort, Ampersand improves sharply on sharing information (65% to 89%), handling conflict (11% to 30%) and emotional intelligence (25% to 40%). But it falls on resisting unsafe pressure from a teammate, from 88% to 68%, and on taking over a compromised session, from 51% to 41%.

Confirmed vs unconfirmed

Claim	Status
Ampersand v0.1 is live via API with no waitlist	Confirmed in docs
Built on GLM 5.3, about 744B total and 40B active parameters	Confirmed in model card
$1.00 input, $0.20 cached, $5.20 output per million tokens	Confirmed on pricing page
Users “accomplished more” than with Fable in simulations	True on goal progress only; tied once humans&’ default cost penalty applies
59.1% teamwork score	Self-reported run; not on the benchmark’s public board
Gains carry over to real human groups	Unproven; the model card says results “may not generalize to real groups”

The safety numbers worth reading twice

The model card’s safety table is more candid than most. Accuracy on WMDP cyber, a proxy for hazardous cybersecurity knowledge, rose from 80.6% to 89.0%, an 8.4-point jump the card itself calls “the exception.” Prompt injection barely moved (0.668 to 0.687 on SecFid), and humans& writes that its successful injection rate is “substantially higher than frontier models.”

There is good news too. The attack risk score on FORTRESS fell from 21.4 to 13.8, and conversations with sycophancy dropped from 77.3% to 62.6%. On coding, Ampersand slipped slightly, resolving 82.4% of SWE-bench Pro v2 Hard-51 tasks against 84.3% for GLM 5.3.

For a model meant to read every message in a group and call tools, weak injection resistance matters. Anyone in the chat can paste text that tries to steer it.

What it means for Indian developers

At about ₹95 to the dollar, Ampersand costs roughly ₹95 per million input tokens and about ₹494 per million output tokens. That is workable for a community or team assistant on WhatsApp-style group flows, Slack or Discord, where the bot reads far more than it writes.

Three things to plan for. Treat every group message as untrusted input and gate tool actions behind human approval, as the model card itself advises. Watch the 10,000 lifetime session cap on the Normal tier if you open one session per chat. And because the model holds memories about people, make its presence clear and get consent; the card asks for this, and India’s Digital Personal Data Protection Act, 2023 puts consent at the centre anyway.

If you want an open model to self-host instead, the base GLM 5.3 weights are available, but Ampersand’s tuned weights are not. Developers comparing cheaper agent models may also want our breakdowns of DeepSeek V4.1 Flash and MiniMax M3.1 Flash, and our look at how OpenAI tests GPT-6.1 Sol.

The bottom line

Ampersand is a credible first release: a fine-tune that lifts an open model to the level of Claude Fable and Gemini 3.8 Flash on humans&’ own social simulation, with clearly published prices and an honest model card. The marketing line runs ahead of the data. It talks too much for some users, trails closed models on independent teamwork scores, and every headline result so far comes from simulated people, not real ones.

FAQ

What is Ampersand v0.1?

It is the first model from the AI lab humans&, released on 20 October 2026. It is a version of Z.ai’s GLM 5.3 trained to work inside group chats and decide when to reply or stay quiet.

How much does the Ampersand API cost?

$1.00 per million input tokens, $0.20 per million cached input tokens and $5.20 per million output tokens, including reasoning tokens. Normal and High tiers share the same prices but have different rate limits.

Does Ampersand beat Claude Fable?

On raw goal progress in humans&’ simulation, yes, 13.82 to 13.45. After the company’s default penalty for unwanted interactions, the scores are 12.943 and 12.945, a statistical tie. On the Stanford teamwork benchmark, Fable 5.1’s public score of 66.6% is above Ampersand’s self-reported 59.1%.

Are Ampersand’s weights open?

No. It is available only through the humans& Agents API. The GLM 5.3 base model it is built on has open weights.

Is Ampersand safe to connect to tools?

Use care. humans& says its prompt injection success rate is substantially higher than frontier models, so keep human confirmation on any consequential action.