Data platform Scale Labs released its latest Audio MC S2S leaderboard today, and the results have a plot twist: a relative newcomer has pulled even with OpenAI’s best real-time speech model. Thinking Machines’ freshly launched TML-Interaction-Small model scored an APR of 43.4%, tying with OpenAI’s GPT-Realtime-2 (xHigh) for the top spot.
https://twitter.com/ScaleAILabs/status/2053974561805496374
In raw numbers, GPT-Realtime-2 (xHigh) earned 48.45 points, while TML-Interaction-Small clocked in at 43.36. Because the gap falls within the statistical margin of error, officials declared it a tie for first place. The next tier includes the standard GPT-Realtime-2 (37.61 points), Google’s Gemini 3.1 Flash Live with thinking mode enabled (36.06 points), and the older GPT-Realtime-1.5.
Scale Labs noted that TML-Interaction-Small maintains fast conversational response times while showing long-context awareness that’s rare among existing full-duplex models.