Skip to the document
Madhuopen lab

AI Insights: Latency & Cost Optimization · part 10 of 17

adopted27 July 2026cost: API-only (no GPU)

GPT-5.1 speed ablation: what actually makes the API faster

Prompt compression, output caps, reasoning_effort, and the Azure TTFT floor

The question

If the paid GPT-5.1 path stays (as fallback or primary), how much faster and cheaper can the identical call get?

adopted

Output length is the only real lever: a terse prompt cut wall time 32% and cost 33% with quality held at 8.0.

Terse prompt (S4)13.9 → 9.5 s−32% wall, −33% cost
Quality held8.0 Opus"exactly 6 concise insights"
Azure TTFT floor~1.2–1.5 ssub-second needs streaming UX or Qwen
Prompt cache2,688 tokalready active on Azure

What worked

Telling the model to write "exactly 6 concise insights" (variant S4) cut wall time from 13.9 s to 9.5 s and cost by a third, with the Opus judge score holding at 8.0. Output tokens are both the latency and the cost — the model writes at a fixed pace and output is priced 8x input, so every word not written is saved twice.

More aggressive compression (S5/S6, ~6.4 s) costs about one judge point — usable as an optional fast tier, not as the default.

What backfired

The reasoning_effort trap on Azure GPT-5.1: "minimal" is a no-op, and "low" actually enables reasoning — TTFT jumped 1.5 s → 13.3 s and cost rose 66%. Never set it. Any reasoning/verbosity parameter also makes the API reject temperature/top_p.

Azure's prompt cache was already active (2,688 tokens cached per call) — confirming the sub-experiment 6 side-finding that keeping the shared prompt identical and in front pays on the paid API too.

GPT-5.1latencyoutput capreasoning_effortprompt cache