Skip to content- Sources: primary
- Summary: In a 2026-11-13 video, Theo says an audit of his own agent sessions found Gemini 4 Carbon consistently about five times slower than other models on comparable work, with over 11 seconds between request acceptance and the start of a response on each step. He reports that it reads a file 8 to 16 times on average, that about 30% of reads miss a cache it just wrote, and that its price doubles on 2026-12-01. He quotes a Google lead saying time to first token is high because prefill is costly and that the team is looking at optimizations.
- Why it matters: Per-step latency compounds across multi-turn agent loops, so measure time to first token inside the harness, not single-request benchmarks, before routing agent work to a model.
- Follow-up: Gemini 4 Carbon price change on 2026-12-01.