Mini Is a Price Tier. It Is Not a Latency Tier.
gpt-4.1-mini costs five times less per output token than gpt-4.1. It is also slower to first token in all four runs we have measured since June, by an average of 105 ms. A smaller model does not reach the first word sooner on the one-sentence answers a voice agent generates.
Our LLM board had gpt-4.1 reaching the first token in 484 ms and gpt-4.1-mini in 592. It looks like two numbers that got swapped on their way into the table.
They didn’t. Every run we have says the same thing.
Mini is slower in all four
Four runs, four weeks apart at the widest, each one measuring both models back to back from the same machine in Virginia on the same prompt. Mini lost every time, by 65 to 178 ms, averaging 105. A stratified permutation test across the three runs where we kept every individual call reading puts the two-sided p at roughly 0.0001. One afternoon of bad luck on a shared serving fleet does not do this four times.
Any single call can go either way
The two clouds sit on top of each other. Pick one call from each model and mini has a real chance of winning it, which is why an eyeball test in a playground will never settle this. The separation lives in the middle of the distribution and in the tail. Across 42 calls per model, gpt-4.1 crossed one second once. Mini did it four times, and its worst call took 5.7 seconds against gpt-4.1’s worst of 3.9.
What the first token is waiting for
The prompt is one sentence and the answer is 31 characters. On a turn that short, time to first token is connection setup, whatever queueing and scheduling the vendor’s fleet applies, and prefill on about twenty prompt tokens. Decode speed barely enters it. A smaller model does generate tokens faster once it is generating, and at 31 characters there is nothing left there to win.
Why the two tiers differ, we are guessing: separate pools, each with its own occupancy, batching and autoscaling. What we can measure is that first-token time tracked the serving path rather than the parameter count. Generation itself was also slower per character for mini in two of the three runs, though at this length that number is too noisy to lean on.
What mini buys is its name. Output tokens cost $1.60 per million against $8.00 for gpt-4.1. Five times cheaper is the reason to pick it.
Where the assumption gets expensive
First-token time is half of how the LLM board ranks, because it is the stretch of the pipeline the caller sits through in silence. “Pick the mini tier, it’s the fast one” costs about 70 ms of dead air per turn. It also buys a model that fabricated on 20% of our out-of-policy questions. gpt-4.1 fabricated on none.
None of this is specific to OpenAI. Every vendor ships a mini, nano, flash, lite or turbo tier, and the name is a claim about cost and capability. Whether it reaches the first token sooner is a question about their infrastructure, and the only way to know is to measure it in the region you serve at the answer length you generate.
The second bug this turned up
The two numbers that started the argument were three weeks old. Both 4.1 rows were measured on 3 July and carried forward untouched while other rows in the same column were refreshed on 24 and 25 July. One column, several vintages, on the metric that carries the most weight in the ranking.
The fresh run moved both: gpt-4.1 from 484 to 640 ms at p50, mini from 592 to 708. Our control model, measured in the same sweep, read 212 ms against its board value of 195, so the vantage was healthy and the change is upstream of us. The order between the two models held. Both rows now carry the date they were measured.
We nearly got the reading wrong. Before the probe ran we simulated refreshing only gpt-4.1, and predicted mini would overtake it. It didn’t, because mini had slowed by about as much. Refreshing one side of a comparison and reading a rank change off it is the same mistake as the staleness we were fixing.
How this was measured
One machine in Virginia, both models in the same sweep, streaming, the same one-sentence prompt every time. Time to first token is the first byte of the answer. One warm-up call per model is thrown away, then 15 measured calls each, with an unrelated third model in every sweep as a control on the vantage. Runs on 30 June, twice on 3 July, and 27 July. One region, one prompt shape, one vendor family. Per-call readings survive for three of the four runs, which is why the second chart shows three.