GLM-5.3: The Fastest-Streaming Endpoints Take 18–26 Seconds to Produce the First Token

When choosing a provider for a model, the first figure you usually check is “tokens per second.” The higher, the better. But a recent Artificial Analysis post about GLM-5.3 shows that this figure can be highly misleading about the actual user experience.

What They Found

The GLM-5.3 endpoints with the fastest streaming produce their first token after 18–26 seconds. With most other providers, the response begins within 1–3 seconds. If you measure only stream speed after generation starts, the former appear to be the leaders. Add the initial wait to the request’s total duration, and they drop down the ranking.

The post does not explain the cause, leaving us to speculate. Possible explanations include a long queue, a cold start, heavy prompt preprocessing, or the model reasoning before returning its answer. Without data from the providers themselves, the culprit cannot be identified with confidence.

Why This Matters in Practice

Imagine a customer-support chatbot. A user sends a question and stares at a blank screen. If the response starts 20 seconds later, it does not matter that the text then flies in at 300 tokens per second. The user may already have closed the tab or clicked “send” again.

The same applies to agents. Every step in a chain adds the delay before the first token. Across ten consecutive calls, the difference between 2 and 20 seconds turns into minutes of waiting rather than a minor detail.

That is why interactive products should track at least three metrics:

• time to first token (TTFT);
• output speed after generation starts;
• total request time for a typical response length.

What to Do

First, do not choose a provider based on a single speed figure from an aggregator. Open the detailed breakdown and inspect TTFT separately.

Second, run your own measurements with your own prompts and at your usual time of day. Provider load fluctuates, while public benchmarks average the results.

Third, if your responses are short, TTFT almost entirely determines how fast the product feels. For long responses, streaming speed matters more, but even then you will need to mask a 20-second delay somehow, for example with a progress indicator.

Fourth, build in a fallback. If the first token has not arrived within 5 seconds, switch to another endpoint.

Limitations

This is a single short post with 807 views, no table in the post itself, and no provider names. We do not know how long these delays persisted or whether they are reproducible. The conclusions should be checked against live Artificial Analysis data and your own tests.

But the lesson itself is universal: “fast streaming” and a “fast response” are not the same thing. What matters to the user is when the first word appears, not how quickly the tokens arrive afterward.