Gemini 3.8 Live Extended Thinking vs GPT Realtime 2.1

GPT Realtime 2.1 is more reliable, by 8.6 points. Gemini 3.8 Live Extended Thinking handled 17 of 70 real callers well on all three runs, and GPT Realtime 2.1 23. A caller only counts when every run of the call scored 80 or more.

For a voice agent, GPT Realtime 2.1, since it handles more real callers well on every run. GPT Realtime 2.1 is ahead on ELO rating, hearing score, task handling score, reliability, calls handled well, calls fully resolved and calls where the agent went quiet, and Gemini 3.8 Live Extended Thinking on turn taking score, speaking score and median latency.

GPT Realtime 2.1, by 11.4 points higher. Gemini 3.8 Live Extended Thinking fully resolved 22.9% of calls and GPT Realtime 2.1 34.3%. A call counts only when every step the caller needed was done.

Yes. Gemini 3.8 Live Extended Thinking starts replying 0.06s sooner at the median. Gemini 3.8 Live Extended Thinking takes 2.10s at the median and 2.60s at p90; GPT Realtime 2.1 takes 2.16s and 2.70s. Both are measured on the call audio, from the caller finishing to the agent starting to reply.

GPT Realtime 2.1, by 5.7 points. The agent went quiet for 10 seconds or more on 17.6% of Gemini 3.8 Live Extended Thinking calls and 11.9% 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 with the same prompt, tools and phone connection. Both took the same 70 recorded real human callers on live calls, three times each. Each call is graded on 25 skills across Hearing, Turn Taking, Task Handling and Speaking, and passes at a score of 80 out of 100.