Does a complex task take tens of minutes? The bottleneck is almost always model generation time × Agent rounds, not the relay layer.
Where the Time Goes
Claude Code / Codex CLI processing a task = dozens to hundreds of API rounds. Measurements across all production requests over 24 hours:
- Model generation accounts for 93% of total time
- Relay forwarding overhead is <1 second/request (7%)
A 30-minute task ≈ 100+ rounds × around 10 seconds per round. To make it faster, either reduce the number of rounds or switch to a faster model.
Model Speed Reference
Measured single-request p50:
| Model | p50 | Best for |
|---|---|---|
gpt-5.4-mini | ~4s | Light tasks, small changes, batch processing |
gpt-5.4 | ~9s | Day-to-day development |
gpt-5.5 | ~11s | Complex reasoning, large refactors |
The fastest and slowest differ by nearly 3×. The same applies on the Claude side: claude-haiku-* is faster than sonnet / opus.
Speed-Up Recommendations
Choose Models by Task
Do not use the heaviest model for simple tasks. For summaries, translation, and minor fixes, try gpt-5.4-mini / claude-haiku-* first.
Lower Reasoning Effort
The higher the reasoning level, the longer the model thinks. A single xhigh round can be several times slower than medium. medium is sufficient for day-to-day development; enable xhigh only for particularly difficult problems.
Codex CLI (~/.codex/config.toml):
model_reasoning_effort = "medium"
When Claude Code uses ChatGPT models:
export CLAUDE_CODE_EFFORT_LEVEL=medium
Control Context Length
The longer the context, the slower the time to first token. Run /clear before starting a new task, and use /compact promptly in long sessions.
FAQ
Q: Is the relay service slow? A: Measured relay overhead is <1 second/request, accounting for 7%. Connecting directly to the official service with the same model will not be noticeably faster.
Q: Why is the same question sometimes fast and sometimes slow?
A: The number of rounds varies, as does official service load. Generation time itself also has a long tail (p99 can exceed 5× p50).
Q: Does streaming make it faster? A: Total duration is the same, but the first token arrives sooner, so it feels faster. CLI tools use streaming by default.
Q: Can switching networks / regions improve speed? A: Basically no. Network round-trip time is only a small fraction of total time; most of it is spent generating at the model provider's data center.