Measured: time to first token, thinking time and output speed for nine open chat models

We timed every chat model on the public endpoint with a fixed prompt: time to the first token, how long each model thinks before the visible answer starts, and output tokens per second. Includes a correction to the figures we first published.

Taylor Hawkes · September 16, 2026

Correction, 17 September 2026. The first version of this post divided every generated token, hidden reasoning included, by the time after the first visible word. That overstated the speed of reasoning models, in one case by more than double, and we published figures of 265 to 390 tokens per second that were wrong. The table below uses the standard definitions, and every page that quoted the old numbers has been updated.

Provider pricing pages tell you the cost per token; almost none tell you how long you will wait. We ran the same prompt against every chat model we serve, three times each, and recorded the median. The numbers below were measured on 2026-09-17.

Model	First token	First answer word	Reasoning tokens	Output speed	Total
`openai/gpt-oss-120b`	0.46 s	1.38 s	36 of 196	45.8 tok/s	4.92 s
`deepseek-ai/DeepSeek-V4-Flash`	0.85 s	20.1 s	931 of 1096	54 tok/s	29.95 s
`deepseek-ai/DeepSeek-V4.1-Flash`	0.5 s	3.45 s	456 of 591	151.6 tok/s	4.29 s
`zai-org/GLM-5.3-Flash`	0.52 s	10.31 s	811 of 957	89.7 tok/s	11.19 s
`meta-llama/Llama-3.3-70B-Instruct-Turbo`	0.62 s	0.62 s	0 of 119	98.2 tok/s	1.93 s
`zai-org/GLM-5.3`	0.41 s	6.22 s	1036 of 1185	170.6 tok/s	6.8 s
`deepseek-ai/DeepSeek-V4-Pro`	0.64 s	1.25 s	96 of 235	121.4 tok/s	2.59 s
`MiniMaxAI/MiniMax-M3`	1.3 s	2.74 s	79 of 250	72 tok/s	6.92 s
`moonshotai/Kimi-K2.7-Code`	0.87 s	37.34 s	0 of 1710	40.7 tok/s	39.77 s
`Qwen/Qwen3-VL-235B-A22B-Instruct`	1.28 s	1.28 s	0 of 147	10.7 tok/s	15.26 s

How to read it

First token is the wait before anything arrives, reasoning included. It is under a second for almost every model here, so the network and the queue are not the problem.
First answer word is what a person in a chat box feels. Reasoning models think first, and how long they think dominates everything else: on this easy prompt it ranged from about a second to more than half a minute.
Output speed counts every generated token, reasoning included, from the first token on.

What stands out

Thinking time, not token speed, decides how fast a model feels. DeepSeek V4 Flash and Kimi K2.7 Code generated at similar speeds to GPT-OSS 120B, but spent hundreds to well over a thousand tokens reasoning about a request for one paragraph. GPT-OSS 120B and DeepSeek V4 Pro thought briefly and answered within a second or two.
That thinking is billed. Reasoning tokens are output tokens. A model with a low price per token that thinks ten times as long is not the cheap option for simple requests.
You can turn it down. reasoning_effort: "low" on GPT-OSS and GLM, and "reasoning": {"enabled": false} on DeepSeek, cut the thinking and the wait; each model page shows the measured effect.
Llama 3.3 70B does not reason, so its first token is its first answer word. For short interactive replies that still counts for something.
GLM 5.3 and DeepSeek V4.1 Flash had the highest raw output speed in this run.

Speech and transcription

Model	Task	Measured
`hexgrad/Kokoro-82M`	Text to speech	300 characters → 0.84 s
`canopylabs/orpheus-3b-0.1-ft`	Text to speech	300 characters → 58.5 s
`openai/whisper-large-v3`	Speech to text	19.5 s of audio → 0.92 s (0.047× real time)
`nvidia/parakeet-tdt-0.6b-v3`	Speech to text	19.5 s of audio → 0.7 s (0.036× real time)
`mistralai/Voxtral-Small-24B-2507`	Speech to text	23.1 s of audio → 2.08 s (0.09× real time)
`mistralai/Voxtral-Mini-3B-2507`	Speech to text	23.1 s of audio → 1.35 s (0.058× real time)

Two things worth knowing before you pick a speech model: Kokoro returns a 300-character paragraph in under a second, and Orpheus takes about a minute for the same text. Orpheus sounds more expressive; it is a pre-rendering tool, not an interactive one. Both transcription models run at roughly 4–5% of real time, so a 20-minute recording comes back in under a minute. These figures are from 2026-09-16 and were not affected by the correction.

Method and caveats

Public endpoint api.inferenceapis.com, HTTPS, streaming on, no warm-up, default reasoning settings.
Prompt: "Write a short paragraph (about 120 words) explaining what an API rate limit is." with max_tokens 2000 so that no model is cut off while thinking.
First token: the first streamed chunk carrying reasoning or content. First answer word: the first chunk with non-empty content. Output speed: completion tokens, as reported in the final usage chunk, divided by the time from the first token to the end.
Three runs each, median reported. One easy prompt is a narrow test: thinking time grows with difficulty, and upstream capacity varies by hour, so expect wide variation. Treat differences under about 30% as noise.

The raw JSON is what the model and comparison pages read: compare any two models.