GLM 5.3 Real Cost: Three Numbers Don’t Match the Model Card

👁 117 reads

The GLM 5.3 model card looks appealing: a million tokens of context, input pricing below ten cents for the Flash version, image and video support. It looks like the cheapest way to get large context available today. So we ran our own tests and calculated what this offering actually costs. It turns out three of the four numbers on the model card are only true under very specific conditions.

Our test: we tested how GLM-5.3 handles writing in Polish separately: GLM-5.3 in Polish: 13-task test · comparison with MiMo, Qwen and Muse Spark

What GLM 5.3 Actually Is

GLM 5.3 is a model from Chinese company Z.ai, successor to the GLM 5 line. On OpenRouter it’s available in two variants, and the difference between them is larger than the name suggests.

The first thing that stands out: the cheaper Flash version has a larger context than the full version and accepts images and video, which the flagship doesn’t accept at all. With nineteen times lower input pricing. This isn’t an arrangement you see often, so we started there.

You Won’t Get 1.31 Million Context from Almost Anyone

GLM 5.3 Flash has twenty-six providers on OpenRouter. We checked each one individually and the 1,310,720 token context window is offered by exactly one: Cloudflare. Twenty-four of the rest give 1,048,576, and one gives 262,144.

The number on the model card is therefore true, it just describes the best provider, not what you’ll get with a normal request. If you’re building something that actually needs 1.3 million tokens, you must force a specific provider. Otherwise you’ll get a million and the request will bounce off the limit at the worst possible moment.

Model Card Price Is the Cheapest Provider, Not Your Price

This is where it gets more interesting. OpenRouter shows the cheapest provider’s rate on the card, but routes requests to whoever is currently available. We calculated the spread for each of the four models.

For GLM 5.3 Flash, the card rate of $0.25 is offered by two providers out of twenty-six. Twenty charge $0.50, which is twice as much. Our requests went to Together and we paid exactly $0.50.

With GLM 5.2 it was worse. The card says $0.88 per million output tokens. Our request went to a provider charging $4.40, which is five times more than the price we based our decision on. This isn’t an error or hidden fee, everything is transparent in the provider data. It’s just that nobody reads the provider data, and the model card shows one number.

Reasoning Cannot Be Disabled

Our standard measurement method starts with disabling reasoning, because otherwise we’re comparing models that think with models that answer immediately. With GLM 5.3 this doesn’t work. The API returns a 400 error and a message that leaves no room for discussion:

Reasoning is mandatory for this endpoint and cannot be disabled.

The Flash version gives the same error. The consequence is financial, not technical: reasoning tokens are billed as regular output tokens, and output is the more expensive side of the price list. You pay for the model’s thinking and you can’t opt out.

We checked how much thinking there is. In three tasks, the share of reasoning tokens in total output was:

• GLM 5.3: 95%, 84%, and 52%
• GLM 5.3 Flash: 96%, 93%, and 68%
• DeepSeek V4 Flash: 89%, 16%, and 90%

In a task where we asked for two sentences in Polish, GLM 5.3 produced 659 output tokens, of which 625 were reasoning. We got two sentences and a bill for twenty times more text than we saw.

Three Non-English Tasks, Four Models

The test was simple and deliberately short, because in such tasks cost differences become most painful. We ran it in Polish (a heavily inflected language, useful as a proxy for non-English writing quality): we asked for two sentences with difficult declension, a summary in exactly three sentences, and a calculation on a token pricing menu.

The most striking number is in the summary row. GLM 5.3 Flash needed 3,783 output tokens and 46.3 seconds for three sentences about leasing, of which 3,526 tokens went to reasoning. DeepSeek V4 Flash returned the same task in 203 tokens and 3.8 seconds. Twelve times faster, at a cost thirty-nine times lower.

In total for three tasks, DeepSeek V4 Flash charged sixteen times less than GLM 5.3 Flash and sixty-one times less than GLM 5.3.

What About Non-English Language Quality?

There’s no surprise here and that’s exactly GLM’s problem. All four models handled the genitive plural of “violin” and the instrumental case of “thousand” correctly. All calculated the bill at $0.646 correctly. All fit the summary into exactly three sentences as requested.

Sentence from GLM 5.3: “In the basement they didn’t find the old violins belonging to grandfather.” Sentence from DeepSeek V4 Flash: “I ran out of strings for my old violins.” Both correct, one cost sixty-one times more.

In tasks where the more expensive model does nothing better, the entire difference comes down to the bill. And that’s the answer to whether it’s worth upgrading from GLM 5.2 to 5.3: in our tasks there was nothing worth paying extra for.

Same Model, Four Times Different Response Time

There’s one more thing worth knowing. GLM 5.3 answered us once in 1.7 seconds and once in 29.7 seconds. This wasn’t a difference in task difficulty, but a difference in provider: the shorter response came from Modal, the longer from DigitalOcean. If you’re building something where response time matters, with models that have long provider lists you must specify a particular one. Otherwise the same code works in one second one time and half a minute another, and you don’t know why.

Who Should Use This

Take GLM 5.3 Flash if you need a cheap model that accepts images and video with large context. This is a genuinely rare combination and Flash wins here, because DeepSeek V4 Flash doesn’t accept images.

Skip the full GLM 5.3 until you have a task that the Flash version fails. It costs almost nineteen times more on input, doesn’t accept images, has smaller context, and in our tasks didn’t answer better.

For plain text, take something else. For text tasks, DeepSeek V4 Flash was faster and many times cheaper with the same language quality.

Always specify the provider. It’s one line in the request, and without it price and response time are a lottery where the model card shows the best possible result.

How We Measured This

All numbers are from September 9, 2026. We took prices and provider lists straight from the OpenRouter API, costs of individual requests from the billing field returned by the same API, not from our own estimates. Each task went to each model once, with the same prompt and a 4,000 token limit. The entire test cost $0.0196.

One measurement is one measurement, and with reasoning models the variance can be significant. So we’re not claiming GLM 5.3 Flash always needs 3,783 tokens for three sentences. We’re claiming that on the first attempt it did, that you can’t disable this, and that you need to calculate bills with that assumption.

We collect more such measurements continuously in benchmarks, and keep data on individual models in the model atlas. For how similar traps look with more expensive models, we detailed it in the article about GPT-6 Astra.

Same-Day Update: Broader Test Came Out Better for GLM

A few hours after publishing this article, we ran ten models through six Polish grammar tasks, scored automatically. GLM 5.3 Flash scored full points, the same as GPT-6 Astra, Claude Fable 5.1, and Kimi K3, with a bill twenty-six to thirty-two times lower than theirs.

This refines the conclusion from the paragraph above. On the three tasks in this article, DeepSeek V4 Flash performed equally well and was cheaper. On a broader set, DeepSeek lost a point and GLM 5.3 Flash didn’t lose any. If you care about correct Polish, GLM 5.3 Flash is now our default choice, and everything we wrote above about reasoning costs and providers still applies.

Full results, broken down by task and with the cost of each model, are in the article about which AI model writes best in Polish.

Second Time the Same, Only More Painful

With longer tasks, mandatory reasoning stops being just a cost. In a test where we asked eight models for a 900-word article, GLM 5.3 Flash used the entire 8,000 token limit on thinking and returned zero words, while still billing us.

Google · Your sourcesQuantslant higher in your Google - one clickAdd to preferred sources →

This Is No Longer an Exception

When we wrote this article, GLM 5.3 was the only model where we encountered this. A day later we reviewed eighteen models and found that seven of them block disabling reasoning, and the division is purely generational: new models take away the choice, older versions of the same families still allow it.

Data: GLM 5.3 card with prices per million tokens and context size.

Kinetyra