GLM-5.3-Flash vs DeepSeek V4 Flash: Neither One Has a Single Price

Both models are called Flash and neither of them has a price you can put in a spreadsheet cell. DeepSeek V4 Flash charges by the hour of day, doubling its rate for 21% of the week. GLM-5.3-Flash’s cheap number is a promotion with an expiry date on it. So “which is cheaper” has no answer until you say when you call, how much of your input hits cache, and how much output you generate.

The useful part is that all three are computable from published rate cards. No benchmark required.

Rates below were read from DeepSeek’s pricing documentation and Z.ai’s pricing page on 2026-08-28. Disclosure: Ofox resells both models, and one fact that reflects badly on us is in the fourth section.

What do the two rate cards actually look like?

Four of them, because each vendor publishes two. US dollars per million tokens.

The two second columns are different kinds of thing, and that distinction matters more than it looks.

GLM’s $0.075 is a promotion. Z.ai’s page says so directly: “GLM-5.3-Flash is available at a 50% discount (strikethrough prices are list prices). The promotion ends at 24:00 on September 9, 2026 (UTC+8, Singapore time).” After that it reverts to $0.15, once and permanently.

DeepSeek’s off-peak rate is not a promotion. It is permanent structure. Peak is defined as “01:00 - 04:00 and 06:00 - 10:00 UTC, Monday through Friday”, everything else is off-peak, and “off-peak rates are half of the peak rates”. That is 35 hours out of 168, so roughly 79% of the week runs at the cheaper number.

Where you are decides how much of that 79% you actually get. In US Eastern the peak window lands at 21:00-00:00 and 02:00-06:00, almost entirely outside a working day. In Japan it lands at 10:00-13:00 and 15:00-19:00, squarely inside one. Same rate card, materially different bills.

Why does “which is cheaper” have no single answer?

Because the two vendors win on opposite billing lines.

Taking GLM list against DeepSeek off-peak, so neither side gets a temporary advantage:

DeepSeek’s cache is dramatically cheaper and its output is dearer. Those are not small differences pointing the same way; they are large differences pointing opposite ways, and which one dominates is a property of your workload rather than of either model.

A long-document classifier that replays the same context and emits twenty tokens lives almost entirely on the cache line. A coding agent that returns several thousand tokens per turn lives on the output line. Same two models, opposite conclusions.

What cache-hit rate does DeepSeek need to win?

This one has an exact answer, because it only depends on published prices.

Let h be the share of input tokens that hit cache, and r the ratio of output tokens to input tokens. DeepSeek off-peak costs less than GLM list when:

0.22(1-h) + 0.007h + 0.66r  <  0.15(1-h) + 0.03h + 0.50r

which reduces to:   h > (0.07 + 0.16r) / 0.093

Substituting:

Once output exceeds roughly 14.4% of input volume, a cache 4.29 times cheaper cannot rescue an output line 1.32 times dearer. The discount only applies to input, and there is no equivalent lever on the output side.

So the selection procedure does not start with a rate card. It starts with two numbers from your own logs: cache-hit rate, and output-to-input token ratio. Most conversational and drafting workloads sit far above 14.4% and can stop reading there. The workloads that genuinely land on DeepSeek’s side are long-context, high-replay, short-output — the shape of an agent loop replaying a large system prompt for a small tool call. We worked a neighbouring case in Gemini 3.1 Flash Lite vs DeepSeek V4 Flash for agent loops.

Why does our own gateway make GLM look better than it is?

Because we do not price the two models the same way, and you should know that before you use our list to decide.

Our DeepSeek V4 Flash model page carries $0.44 input, $1.32 output and $0.014 cache read. Compare those three to the table above: they are DeepSeek’s peak rates, exactly. A request through our gateway pays peak around the clock and never sees the off-peak half-price window that covers 79% of the week.

GLM we handle differently. Our GLM-5.3-Flash page lists $0.075, $0.25 and $0.015 with list prices struck through beside them, which is Z.ai’s promotion passed through unchanged. When that promotion ends on 2026-09-09, our page follows it back up.

The consequence is straightforward and it is not in our favour: comparing these two models on our own price list overstates GLM’s advantage. If you are using gateway numbers to decide, halve the DeepSeek column first and look again.

What differs besides price?

Three things, all documented.

Output ceiling, by a factor of 2.9. DeepSeek V4 Flash tops out at 384K output tokens; GLM-5.3-Flash at 131,072. Context is 1M on both sides, so the difference is not what you can feed in, it is what can come back in one call. Generating a long report or a large batch of structured records will hit 131K before it hits anything else.

Protocols and concurrency. DeepSeek publishes both an OpenAI-format endpoint at https://api.deepseek.com and an Anthropic-format one at https://api.deepseek.com/anthropic, documents a 2500 concurrency limit for the flash tier, and defaults to thinking mode with a non-thinking mode available. If you are moving a toolchain built against Anthropic’s shape, that second endpoint saves real work.

Weights and architecture. GLM-5.3-Flash is open-weight, 320B total parameters with 18B activated across 45 layers, which Z.ai describes as the first open-source frontier model using a hybrid of sparse and linear attention. What each of those numbers actually governs is in GLM-5.3-Flash’s three parameter counts. DeepSeek’s API documentation says nothing about weight availability for V4 Flash, so we make no claim either way.

How should you decide?

Four steps, ordered by how much money each one moves.

1. Pull two numbers from your logs before opening either rate card: cache-hit rate on input, and output tokens as a share of input tokens. The crossover table above then answers the question directly.
2. If output exceeds 14.4% of input, choose GLM and stop. DeepSeek’s cache advantage is mathematically unable to close that gap, so there is nothing left to model.
3. For long-context, high-replay, short-output work, price DeepSeek direct rather than through a gateway. And compute your real off-peak share instead of assuming 79%, because that figure collapses if your traffic tracks a working day in a timezone where peak lands mid-morning.
4. Put 2026-09-09 in the calendar. GLM’s unit cost doubles back to list that day. At list it still leads on uncached input and output, so most conclusions survive, but confirm yours rather than inheriting it.

Endpoint and parameter details for the full-size sibling are in the GLM 5.3 API guide, and the DeepSeek-side cost levers are in paying less for DeepSeek V4 Flash.

Both vendors published the rules in full, down to the clock hours of the peak window. The work left is not guessing which is cheaper. It is measuring what your own workload looks like.

References

• DeepSeek models and pricing
• DeepSeek release notes
• Z.ai pricing overview
• Z.ai GLM-5.3-Flash model documentation
• Ofox DeepSeek V4 Flash model page
• Ofox GLM-5.3-Flash model page