Gemini 3.8 Live Extended Thinking vs GPT Realtime 2.1

GPT Realtime 2.1 is more reliable, by 26.8 points. Gemini 3.8 Live Extended Thinking passed 43 of 82 scenarios on all three runs, and GPT Realtime 2.1 passed 65. A scenario only counts when every run of it passed.

For a voice agent, GPT Realtime 2.1, since it gets more scenarios right on every run. GPT Realtime 2.1 is ahead of Gemini 3.8 Live Extended Thinking on reliability, success rate, agent responses, data accuracy, calls where the agent went quiet, interruption and response time.

GPT Realtime 2.1, by 7.6 points. Gemini 3.8 Live Extended Thinking saved everything exactly right on 84.4% of the calls that save data, and GPT Realtime 2.1 on 92.0%. On Booking calls, 88.9% against 95.9%. On Medicare calls, 72.7% against 81.8%. A call counts only when nothing is missing, wrong, extra or saved twice.

No. GPT Realtime 2.1 starts replying 0.13s sooner at the median. Gemini 3.8 Live Extended Thinking takes 2.07s at the median and 2.74s at p90; GPT Realtime 2.1 takes 1.94s and 2.46s. Both are measured on the call audio, from the caller finishing to the agent starting to reply.

GPT Realtime 2.1, by 0.03 on a 0 to 5 scale. Gemini 3.8 Live Extended Thinking scores 4.95 and GPT Realtime 2.1 4.98 for how well the agent handled being talked over.

GPT Realtime 2.1, by 2.0 points. The agent went quiet for 10 seconds or more on 5.3% of Gemini 3.8 Live Extended Thinking calls and 3.3% of GPT Realtime 2.1 calls.

There is no verified per-minute rate for Gemini 3.8 Live Extended Thinking, so cost cannot be compared yet.

Cekura ran both as the whole agent in the same open-source Pipecat pipeline, with the same prompt, tools and connection. A simulated caller worked through all 82 scenarios on live calls, three times each, 3,690 calls across the benchmark. A call passes only when the agent said the right things and saved the right data.