The cache-read catch: why GLM 5.3 costs more in real agent work

We went through real bills for agent work and saw a simple thing. For GLM 5.3, the lion's share of the cost is not the answers and not the fresh input — it is re-reading the cache: tokens the model has already seen within the same job. In our measurements that is 80–90% of the bill, and inside long tasks it approaches 98%. That is why the same job consistently comes out several times cheaper on DeepSeek, even though GLM's price list looks modest at first glance. Here is how that happens.

Where the cache in your bill comes from

Every API call happens without memory: the model receives the whole conversation again — system instructions, chat history, file contents. Providers cache the repeated part: if the beginning of a request matches something already processed, those tokens are read from the cache at a discount. The API response shows the breakdown: how many tokens came from the cache, how many were new, how many the model generated.

Sounds like pure profit. But the discount is a rate, not a gift, and every model sets its own. The gap between models is measured not in percent but in multiples. That is where the catch hides.

What the price list says

Public prices as of late August 2026, dollars per million tokens:

Look at the middle column — the one usually skimmed past. For DeepSeek V4 Pro, cache reads cost three percent of the input price: an almost thirtyfold discount for re-reading. For GLM 5.3 it is nineteen percent, a fivefold discount. The Flash models are closer in relative terms (about 20% versus 23%), yet in absolute numbers GLM Flash cache reads still cost roughly twice as much as DeepSeek Flash's.

Measured on real work

We took two weeks of real agent traffic in August: 13,279 API calls across four models — GLM 5.3, GLM 5.3 Flash, DeepSeek V4 Pro and DeepSeek V4 Flash. From those we assembled almost three hundred full task dialogs, where a model works on one job for a long stretch: reading code, editing files, keeping the whole history in context. The typical picture of such a task: 97–98% of input tokens are cache hits. Only a trickle of genuinely new text comes in; the bulk is that same re-read memory.

And here is how the bill breaks down:

• GLM 5.3: 80–90% of the total goes to cache reads — 98% inside a median task dialog. Answers and fresh input are the leftovers.
• DeepSeek V4 Pro: about two thirds of the bill (64%) is also cache. A similar share in spirit, but the absolute amounts are incomparable, because DeepSeek's read rate is twelve times lower.

Now the money, per whole task rather than per call. A median task dialog cost about six cents on GLM 5.3 Flash and about two cents on DeepSeek V4 Flash. To compare fairly, we matched tasks with the same volume of generated text: on comparable workloads DeepSeek comes out 3.6–5 times cheaper. Full GLM 5.3 against DeepSeek V4 Pro, on the class of tasks where they overlap, was 6–7 times more expensive — for the same amount of work done.

A methodology note: these figures are recalculated from the public rates in the table above, not someone's actual invoice. As a sanity check we compared the recalculation with recorded charges wherever the tariff is open — they diverged by 2–3%, so the estimate is sound.

Want to feel it on a single call? A typical agent call in our measurements is around 130 thousand cache tokens, three thousand new input tokens and a couple hundred output. GLM 5.3 charges about four cents for such a call. DeepSeek V4 Pro — about half a cent. Seven times, for the same volume of work.

So is GLM cheating?

Formally, no: all rates are published, and the recalculation matches actual charges wherever the tariff is open. The trick lies elsewhere. The eye catches the input and output prices, while the "cache read" line looks like a technical footnote. In a two-message chat it is one. But in agent work, where memory is re-read hundreds of times per task, that line becomes the main cost item. GLM priced it twelve times higher than DeepSeek did for the flagship models — and on long sessions it outweighs everything else. Even DeepSeek's cache is not free; its rates are simply so small that re-reading the whole memory costs pennies.

What to do about it

Three things worth doing before you pick a model for agents:

• Find the cache-read rate in your provider's price list (cache read / cached input). No such line? Ask directly. For long sessions it matters more than the input price.
• Look at your own load profile: the API response shows how many tokens came from the cache. Above 90% cache hits, you are mostly paying for re-reading, not for work.
• Cut memory ruthlessly. Old messages, giant system instructions and just-in-case files are re-read at your expense on every call. Sometimes a fresh session is cheaper than a three-day history.

Both models are available in the NeuralSpace chat, in the Code section and through the public API — run your own task on both and compare the bills. Just remember to open the breakdown: the interesting part always lives in the cached-tokens line.