Speed index

How fast each model answers, measured from Indonesia.

Every hour we send each model the same short test request from Indonesia, through the same public address you use, and record how fast it answers. These are the results, failures included.

Loading measurements…

Speed by model, last 24 hours
ModelFirst token · p50 lower is fasterFirst token · p95 Tokens / second higher is fasterSuccess rate Tests
DeepSeek V4.1 Flashdeepseek-v4.1-flash
DeepSeek V4 Prodeepseek-v4-pro
GLM-5.3 Flashglm-5.3-flash
GLM-5.3glm-5.3
  • DeepSeek V4.1 Flashdeepseek-v4.1-flash
  • DeepSeek V4 Prodeepseek-v4-pro
  • GLM-5.3 Flashglm-5.3-flash
  • GLM-5.3glm-5.3

p50 is the median: half of the tests were faster. p95: 95 of every 100 tests were faster. Tests is the number of test requests sent in the window.

Method

How we measure.

Enough detail to judge the numbers, and to know what they do not tell you.

  1. One test per model, every hour

    An automated job sends each model one streaming request an hour: the same short prompt every time, with at most 64 output tokens.

  2. Measured like a customer

    The tests are sent from a server in Indonesia to the same public address you call. The clock starts when the request is sent, so the figures include our edge and gateway, but not your own network. If our service is down, the test counts as a failure.

  3. Our own requests only

    Every number on this page comes from those test requests. No customer traffic is measured or used.

  4. Your results will differ

    Speed depends on prompt size: a long prompt or a long answer takes longer than this short test. Read the figures as a comparison between models and over time, not as a promise for your workload.

What each number means

Time to first token
From the request being sent to the first token the model streams back, in milliseconds. For a model that always reasons first, that is its first reasoning token. Shown as p50 (the median) and p95.
Tokens per second
Output tokens divided by generation time. Shown as p50.
Success rate
Tests that returned a complete answer, out of all tests sent. A test that fails for any reason, ours included, counts as a failure.
Tests
How many test requests a figure is based on. The fewer the tests, the less certain the figure.
Where the data lives
Each result is written to an analytics database. This page reads the aggregates and refreshes them about every 10 minutes; days and hours are in WIB (Asia/Jakarta).

Send your first request today.

Top up from $1. No subscription, no expiry.

Questions? Our Indonesian support team is at [email protected].