Ampersand: Humans& model focused on group conversation

In this article you will see:

• What Ampersand is and who is behind it
• How the model was trained and what sets it apart
• How to access it via Vercel AI Gateway
• What to expect in upcoming releases

Who is Humans&

Humans& was founded by Eric Zelikman, former researcher at xAI, together with Andi Peng, former DeepMind and Anthropic, and Yuchen He, former OpenAI and xAI. The company focuses on models that facilitate collaboration between humans and AI systems, instead of models that only respond to individual prompts.

Ampersand is the first result of that focus. The name reflects the idea of "and", connecting people and agents in the same conversation.

What sets Ampersand apart

The model was trained with a focus on two capabilities. The first is long term coordination, which allows the model to maintain context and objectives throughout extended conversations with multiple participants. The second is chain reasoning, which allows the model to decompose complex tasks into steps.

Humans& claims that Ampersand achieves state of the art results on group conversation benchmarks, although detailed numbers have not yet been published independently. The model was also trained to work well with tools and APIs, which makes it a candidate for use in agents.

Access via Vercel AI Gateway

Ampersand is available on Vercel AI Gateway with identifier humans-and/ampersand-1. This means developers who already use the gateway can try the model without changing infrastructure.

Integration with AI Gateway also offers observability, usage limits and fallback to other models. For those who already have a flow with multiple models, adding Ampersand is a configuration change.

Benchmark and independent validation

Humans& claims that Ampersand achieves state of the art results on group conversation benchmarks. However, detailed numbers have not yet been published independently. This means the claim should be treated as marketing until there is third party validation.

For those evaluating models, it is recommended to test on real tasks before migrating production workloads. Group conversation benchmarks are difficult to reproduce, and most existing suites focus on individual tasks.

Comparison with existing models

Ampersand enters a market already occupied by models like Claude, GPT-5 and Gemini, which also support multi turn conversations and tool use. The proposed difference is the explicit focus on coordination between multiple participants, humans and agents.

If this difference is confirmed in real tests, Ampersand could be useful in scenarios like discussion rooms with agents, collaborative code review and customer support with multiple specialists. If it is not confirmed, it will be one more model among others on the gateway.

Context: the wave of specialized models

The launch of Ampersand reflects a 2026 trend: instead of one general model for everything, companies launch models optimized for specific tasks. Ampersand focuses on group conversation. Other models focus on code, reasoning or vision.

This specialization has advantages and risks. The advantage is superior performance in the niche. The risk is fragmentation: each task may require a different model, and orchestration between them becomes the central challenge.

What to expect next

Humans& describes Ampersand as the first in a series of models focused on collaboration. The company signaled that future releases will deepen integration with tools and the ability to coordinate between agents.

There is still no date for the next model. The company has also not published its own pricing; usage cost depends on what Vercel AI Gateway charges for the model.

Limits and considerations

As Ampersand is a newly released model, independent benchmark coverage is limited. Before migrating production workloads, it is worth testing on real tasks and comparing with the models you already use.

It is also important to note that the focus on group conversation may not translate into better performance on individual coding tasks. The model is a candidate to test, not an automatic replacement.

How to test today

1. Access Vercel AI Gateway.
2. Use the identifier humans-and/ampersand-1.
3. Test in conversation scenarios with multiple participants.
4. Compare with the models you already use on real tasks.
5. Check Vercel's changelog page for updates.

If you already coordinate agents in projects, it is worth comparing this approach with agent orchestration on bridgecode and with multi agent workflows.