Mira Murati's Thinking Machines announces full-duplex AI model with 0.40-second response time

Thinking Machines Lab, the AI startup founded by former OpenAI CTO Mira Murati, announced what it calls "interaction models" — AI systems that process input and generate responses simultaneously rather than in traditional turn-taking fashion.

The company's first model, TML-Interaction-Small, claims a 0.40-second response time, which Thinking Machines says matches natural human conversation speed and is "significantly faster than comparable models from OpenAI and Google." The technical architecture enables full-duplex communication, meaning the model can listen while it speaks, similar to a phone call rather than a text message exchange.

Release timeline and availability

This is a research preview, not a public product. Thinking Machines plans a "limited research preview" within the next few months, with wider release scheduled for later in 2026. No pricing information has been disclosed.

The company has not released benchmark scores, parameter count, context window specifications, or other technical details beyond the response latency claim.

Technical approach

Current AI models follow a strict turn-taking protocol: user input is processed completely before response generation begins. Thinking Machines' approach processes incoming audio or text while simultaneously generating output, which the company argues should be "native to a model, not bolted on."

The distinction matters for applications requiring real-time interaction, such as voice assistants or conversational interfaces where interruption and natural flow are expected.

What this means

Full-duplex communication represents a meaningful architectural shift if the claimed performance holds under real-world conditions. A 0.40-second response time would indeed approach human conversation norms (typical human response latency ranges from 0.2 to 0.6 seconds). However, without independent verification, public testing, or detailed benchmarks, it's impossible to assess whether this translates to genuinely better user experience or represents a marginal improvement over existing streaming architectures. The value proposition depends entirely on whether simultaneous processing creates noticeably more natural interactions than current streaming implementations, which remains unproven until researchers and users can test the system directly.