Is my LLM slow?

The same prompt, sent to every frontier model, every hour. This tracks how long each one takes to answer.

Loading…

Checking…

Response time

How long each model took to answer, run by run. Switch to vs. its own normal to see each model as a percentage of its own median, which is how the verdict above is decided. A hollow dot means the response came back empty or was refused.

Model health

Each model's latest response time, compared with its own recent runs. Typical is the median of that model's earlier runs in the selected range.

Anatomy of a run

The latest run from each model, split into time spent waiting for the first word and time spent writing.

Waiting for first word Writing (model colour)

How this works

Every model gets the same prompt on a schedule, and we time the reply. Nothing is executed and nothing is graded.

  • One fixed prompt. A hash of it (—) is stored on every run, so any edit to it shows up in the data.
  • Settings are pinned. Reasoning effort and thinking mode are fixed per model and recorded on each run. Both affect latency.
  • Nothing is smoothed. No averaging, no adjusting for length. A model that wrote more took longer.
What the clock covers network provider queue ▼ first byte back — measured reasoning (silent) writing transport thinking — silent total time — the headline clock starts first visible word clock stops

The clock is split three ways. Time up to the first byte of the response covers network round-trip, TLS, and any queueing before the provider starts replying. From there to the first visible word, the model is thinking. After that it is writing.

Network transit and provider queueing cannot be separated inside that first segment, since both finish before the server sends anything. From a single region it stays a small and fairly stable slice.

What each number means

Total time
Request sent to response finished. Includes however much the model chose to write. We do not adjust for length.
First token
How long until the first visible word. This includes reasoning time, so a model that thinks before answering stays silent for all of it.
What we don't measure
Correctness, quality, or capability. Nothing runs the code. The health flag separates a slow run from a refused or empty one. It is not a grade.

All models

Medians across the selected range. Two caveats: rate is not generation speed for models that reason silently and then send the answer at once, and output tokens are not comparable between Opus 4.6 and 4.7+ (different tokenizers). Every model runs at medium effort with adaptive thinking.