DeepSeek V4.1 Flash vs GLM 5.3 Flash: same answer, 12x the tokens

For everyday work, GLM 5.3 Flash costs less. On the same prompts it writes answers about as long as DeepSeek V4.1 Flash's, but DeepSeek spends far more output tokens getting there, and every one of them is billed.

Both are low-cost "Flash" reasoning models with context windows around a million tokens, so on a catalog page the only obvious difference is price. We ran both on the 43 prompts from FastMetal's model comparison set and recorded output tokens and response time. We measured cost, speed and answer length only; we didn't grade answer quality, which depends on the task. The answers themselves are on the comparison and model pages.

Results on 43 prompts

The table averages the 39 prompts both models answered, with both on their default reasoning settings. The run was on September 10, 2026; output cost uses today's USD prices ($0.633 per million output tokens for DeepSeek V4.1 Flash, $0.5275 for GLM 5.3 Flash).

The visible answers are nearly the same length (314 words against 297), but DeepSeek bills 1.7x the output tokens, and with its higher output rate that's about twice the output cost. Response times were a tie.

It isn't every prompt. DeepSeek used more tokens on 25 of the 39 and fewer on the other 14. A few extreme cases pull the average up.

The extreme case: SQL optimization

One of the widest gaps was the SQL query optimization prompt.

The answers are 451 and 443 words. DeepSeek still used 12.3x the tokens. The entire difference is thinking tokens you never see.

Does changing the reasoning setting help?

DeepSeek can turn reasoning off and GLM 5.3 Flash can't, which looks like DeepSeek's trump card. We tested it on the same SQL prompt (costs per 1,000 calls at today's prices):

Two things stand out.

DeepSeek's low doesn't mean "think a little." It still spent 3,396 tokens thinking, hit our 4,000-token cap, and the answer was cut off mid-way. Set low to save money and you may get a truncated answer instead.

GLM 5.3 Flash can't disable reasoning, but at low it used zero thinking tokens and answered in 5 seconds. On this task, at least, not being able to turn it off cost nothing.

So even with DeepSeek's reasoning fully off, its output cost is $1.06 per 1,000 calls against GLM's $0.38, still close to 3x. The trump card isn't one.

Structured output: no longer a difference

When we ran this, the GLM 5.3 Flash route we served didn't guarantee structured output, so DeepSeek V4.1 Flash was the pick for strict JSON schema responses. That has changed. GLM 5.3 Flash moved to a different upstream host in early October 2026, and we verified a strict json_schema call there before marking it supported. Both models now take response_format with a strict schema. DeepSeek V4.1 Flash has moved hosts since the run too, and the token counts above were measured before either move. Reasoning behavior can shift a little from host to host, so check against your own prompts.

Which one to use

• Default to GLM 5.3 Flash. It returns the same amount of answer for far fewer tokens, and its context window is the longer of the two (1.31 million tokens).
• Choose DeepSeek V4.1 Flash if it answers your prompts better. That's a quality question this test didn't answer, so try it on your own work.
• Whichever you use, set reasoning explicitly. Left alone, DeepSeek thinks at high and GLM at max. How to set it is in reasoning control.

Current prices are on pricing and models. Specs and real answers are on the DeepSeek V4.1 Flash and GLM 5.3 Flash pages. The API model ids are deepseek-v4.1-flash and glm-5.3-flash.