TEHNOLOGIE
GLM-5.3-Flash vs DeepSeek-V4-Flash — same tests, four disagreements out of ten
1. What we put head to head
Two budget-tier language models, served by the same infrastructure provider, with the same context window and the same numeric precision. We name them, because otherwise nobody can check the comparison.
| Attribute | zai-org/GLM-5.3-Flash | deepseek-ai/DeepSeek-V4-Flash |
|---|---|---|
| Maker | Z.ai | DeepSeek |
| List price, input, per million tokens | $0.15 | $0.09 |
| List price, output, per million tokens | $0.50 | $0.18 |
| Promotional discount at time of measurement | 50%, applied automatically to the bill | none |
| Effective price paid, input / output | $0.075 / $0.25 | $0.09 / $0.18 |
| Context window | 1,048,576 tokens | 1,048,576 tokens |
| Serving precision | fp8 | fp8 |
On the label, the first one's output costs 2.8 times more than the other's. On the bill, on the day of measurement, only 1.4 times — because one of them carries a 50% promotion applied automatically. The question this article asks is whether either difference shows up in quality.
2. The protocol, briefly
Six tests, all scored automatically, none of them by a third model handing out grades. An automated judge has preferences of its own — the order in which you show it the answers changes its verdict, it favours long answers, and it favours text resembling its own — and it was not needed here: every test has a criterion you can verify by counting.
- Classification. Ten news headlines, two runs each, temperature zero, a single category demanded as the answer. We measure latency, run-to-run stability and agreement between the models.
- Instruction following. Five general-knowledge questions with an explicit 170-word limit and a ban on HTML, links and bulleted lists. We count words and violations.
- Strict format. Five snippets to classify into a JSON object with four exact fields. We check whether the answer parses directly and whether it carries exactly the requested keys.
- Verifiable answers. Twelve arithmetic and elementary-logic problems with a single correct result, searched for automatically at the end of the answer.
- Real cost. The billable tokens each model consumed on exactly the same set of tasks, multiplied by both the list price and the price actually billed.
- Analysis and forecast. An analysis of at most 160 words on the international situation, plus a forecast carrying a claim, a deadline, a probability and a source to check it against, in eight runs per model. The words are counted, the four fields are checked, and the deadline is compared against the day it was written.
The first five tests came to 104 calls, sent to the same endpoint within a one-hour window on 2 September 2026; the sixth added 16 calls the same day, with the same parameters. What follows describes that window, not a permanent property of the models.
3. The results
| Measurement | GLM-5.3-Flash | DeepSeek-V4-Flash |
|---|---|---|
| Classification — median latency | 0.59 s | 0.69 s |
| Classification — 90th-percentile latency | 1.60 s | 2.92 s |
| Classification — same answer on both runs | 9 of 10 | 10 of 10 |
| 170-word limit — breaches | 0 of 10 | 0 of 10 |
| 170-word limit — median latency | 9.13 s | 9.58 s |
| Strict JSON — exact schema | 10 of 10 | 10 of 10 |
| Strict JSON — median latency | 0.97 s | 2.61 s |
| Arithmetic — correct answers | 12 of 12 | 12 of 12 |
| Arithmetic — total output tokens | 548 | 1,462 |
On every test with a known right answer, the two models are impossible to separate: zero length-limit breaches for both, ten valid JSON objects out of ten for both, twelve correct sums out of twelve for both. Neither wrote HTML, links or lists where it had been forbidden to. Had we stopped measuring here, the conclusion would have been that it does not matter which one you pick.
The interesting part starts with the test that has no right answer. On the ten news headlines, the two models returned the same category six times out of ten. The four disagreements are not mistakes:
| Headline | GLM-5.3-Flash | DeepSeek-V4-Flash |
|---|---|---|
| Central bank holds its policy rate and warns about the deficit | economy | finance |
| Annual inflation rises to 5.8% in August | economy | finance |
| A cyberattack takes a bank's services down for six hours | technology / finance | finance |
| The European Parliament votes on amendments to the artificial-intelligence regulation | technology | politics |
A parliamentary vote about artificial intelligence is both politics and technology. Inflation is both economics and finance. A cyberattack on a bank is both things at once — and that is the only headline where one model changed its own mind between the two runs, even though the temperature was zero. There is no answer key to appeal to. A leaderboard awarding points for "correct classification" would in fact be measuring how closely the model resembles whoever wrote the key.
One practical detail from the same test: DeepSeek-V4-Flash returned the same category for all ten headlines on both runs, but once it wrote that category with a capital letter. A program comparing strings exactly would have counted that as a different answer. The model was consistent; the formatting was not.
4. A sixth test: a short analysis and a forecast
The first five tests either have a known correct answer or a formal rule that can be checked by counting. The work these models are actually bought for does not look like that: you read a situation, you summarise it, and you commit to a claim that events can later contradict. The sixth test asks for exactly that — an analysis of at most 160 words on the international situation, followed by a single forecast written in a rigid format: the claim, the deadline, the probability as a percentage, and the public source where anyone can check it. Eight runs per model, same prompt, same temperature.
That format can be graded entirely automatically. You count the words, you check that the four fields are present, you read the probability as an integer — and, the part that matters, you compare the deadline against the date the forecast was written.
| Measurement | GLM-5.3-Flash | DeepSeek-V4-Flash |
|---|---|---|
| Complete format — four fields, valid probability, source | 8 of 8 | 8 of 8 |
| Deadline still in the future on the day of writing | 1 of 8 | 0 of 8 |
| States explicitly that it has no current data | 6 of 8 | 1 of 8 |
| Length of the analysis — median, 160-word limit | 116 | 76 |
| Median latency | 8.7 s | 9.7 s |
| Slowest of the eight runs | 24.7 s | 17.4 s |
| Output tokens, total across eight runs | 2,618 | 2,298 |
| Probability emitted | 70% in seven runs of eight | between 65% and 85% |
Both models passed the format perfectly: sixteen answers out of sixteen with all four fields, a valid probability and a plausible source to check against. An automatic grader that looks only at form would give both of them full marks and close the file.
And both missed the one thing that makes a forecast a forecast. Fifteen of the sixteen forecasts had a deadline that had already passed on the day they were written. GLM-5.3-Flash landed in the future once, with 31 December 2026; its other seven horizons fell on 30 June 2026 or earlier. DeepSeek-V4-Flash never landed: it gave 31 December 2024, 1 June 2025, 30 June 2025 and 31 December 2025 — between eight and twenty months in the past. A forecast whose deadline sits in the past can never be scored, because the answer is already known the moment you read it.
The difference between the two models is not in the form but in what they admit about themselves. GLM-5.3-Flash wrote explicitly that its information was not current in six runs out of eight, in wording that leaves no room for interpretation: "My information network has a time limit, so I cannot confirm the most recent events". DeepSeek-V4-Flash did so once. Otherwise, the closest it came was "There is no recent data indicating a major change in these trends" — a sentence that reads as a claim about the world, not as a limit of the model, and that sounds equally confident whether the world really did stand still or the model simply stopped hearing about it.
The difference also shows in how far each one goes into detail. GLM-5.3-Flash stayed at the level of trends — a war of attrition, technological competition, budget pressures — and named no recent event at all. DeepSeek-V4-Flash did name them, and got them wrong: "the escalation between Israel and Hezbollah risks drawing in regional actors, though Iran is so far avoiding direct involvement" describes a world in which the direct Israel-Iran war of June 2025 never happened. In another run it asked that a forecast be checked against the territories "already controlled on 1 June 2024" — a baseline more than two years old, treated as the present.
One last detail, about calibration: GLM-5.3-Flash answered "70%" in seven of the eight runs, on claims that differed from one another. A probability that does not move with the claim is not an estimate, it is an ornament. DeepSeek-V4-Flash varied between 65% and 85%, which looks better — but since all of its claims were already settled at the time of writing, those numbers cannot be scored either.
This test also overturns the cost arithmetic. Here the model with the higher list price wrote more, not less: 116 median words against 76, and 14% more output tokens. At the prices actually billed, the eight runs cost 1.4 times more with GLM-5.3-Flash, not 1.1 times as in the first five tests. The cost ratio is not a property of the model but of the task you hand it.
5. Real cost is not the list price
The label says 2.8. The bill, on the same set of tasks, says something else — because the pricier model was also the more concise one.
| Across the 52 identical calls | GLM-5.3-Flash | DeepSeek-V4-Flash |
|---|---|---|
| Input tokens | 4,615 | 4,440 |
| Output tokens | 4,504 | 5,130 |
| Total cost, at list price | $0.00294 | $0.00132 |
| Total cost, at the price actually billed | $0.00147 | $0.00132 |
| Cost ratio, measured | 2.23x at list price · 1.11x at the billed price | |
The expensive model emitted 12% fewer output tokens than the cheap one, despite answering the same questions. The difference comes almost entirely from the maths: 548 tokens against 1,462, meaning 2.7 times less talking for twelve answers that were equally correct. On that test, the premium paid drops from 2.8 times to 1.2 times. On the free-response test, where both models fill whatever word budget you give them, the ratio returns to 2.8.
On top of that sits the promotion. At the time of measurement the expensive model was at half price, through a discount applied automatically to the bill, while the cheap one carried none. At the prices actually paid, the 52 calls cost $0.00147 against $0.00132 — a difference of 11%, not 123%. A promotion is not a property of the model, though: it ends, and then the bill doubles overnight without anything in the model having changed.
The operational conclusion is that the price on the label misleads twice over. Once because the ratio on the label applies only to tasks where you decide the length of the answer — where the model decides, verbosity matters as much as the unit price. And a second time because the label you read may not be the price you pay, either today or three months from now. Both are found only by running it and looking at the bill.
6. The mechanism behind a trap that appears in no table
GLM-5.3-Flash deliberates adaptively: on simple prompts it writes no thinking tokens at all, and on prompts it finds ambiguous it writes out a chain of reasoning. The maker exposes a parameter controlling how much deliberation is allowed. The value that ought to mean "none" does the exact opposite of what it promises.
| Deliberation effort requested | Output tokens | What arrived in the answer body |
|---|---|---|
| minimal | 5 | the requested category |
| low | 5 | the requested category |
| parameter not sent | 5 | the requested category |
| none | 300 | the deliberation, as plain text |
Asking the model not to deliberate does not make the deliberation disappear: it moves it out of the field reserved for it and straight into the body of the answer. Instead of one word, 1,485 characters arrive, opening with "The user wants me to classify a news title". The alternative switch, which explicitly disables thinking mode, behaves identically. A three-hundred-token answer instead of a five-token one costs sixty times more and passes any check that only looks at whether the answer is empty.
This is precisely the kind of behaviour no leaderboard reports, because it is not a property of the model but of how the model is called. You find it by running it, on the day you put it to work.
7. An honest verdict
On the four tests with a known right answer, the two models are equal. Not "nearly equal" — identical: 10 out of 10, 10 out of 10, 12 out of 12, zero formatting breaches, for both. The 2.8-fold list-price difference bought, in this measurement, not one point of quality.
What it did buy is speed. GLM-5.3-Flash was faster at the median on all four tests, and on strict JSON it was 2.7 times faster — 0.97 seconds against 2.61. In a service where a person is waiting for the answer, that is a difference somebody feels. In an overnight batch job, it is nothing.
What DeepSeek-V4-Flash bought is steadiness and the bill: the same answer for all ten headlines across both runs, at less than half the list cost. In exchange, its latency tail is longer on classification — 2.92 seconds at the 90th percentile against 1.60 — while the faster model has an unpleasant tail of its own on the maths, where a single call took 21.8 seconds because it decided to think longer.
The sixth test, though, changes the tone of the verdict. Where there is no correct answer laid down in advance, only a common-sense requirement — that a forecast's deadline should sit in the future — both fail, and the cheaper one fails harder: not one valid horizon out of eight, and a single admission that it has no current data. A model that keeps the format perfectly and gets the substance wrong is more dangerous than one that refuses the task, because it passes every automatic check.
So: if you are paying for an answer a person is waiting on, the pricier model earns its money on latency, not on quality. If you are running large volumes with nobody at the other end, the cheap one does the same job at 45% of the list cost — but only 11% cheaper while the promotion lasts, so the sum has to be redone the day it ends. And if what you need is a correct label on an ambiguous piece of text, neither can give you certainty — because here, correctness does not exist as a property of the text. And if you ask either of them for a forecast, check the deadline yourself before using it: both write it convincingly, and both write it wrong.
8. What would invalidate this comparison
- The measurement window. One hour of one day. The latency of a shared service depends on how many others are using it at that moment; a rerun at another hour could reverse the ordering on speed.
- The number of runs. 104 calls, not thousands. The small differences here are not statistically significant and should not be read as if they were.
- The choice of tests. Five short tasks, in Romanian, with verifiable answers. They say nothing about long-form writing, about code or about multi-turn dialogue — the ordering may differ there.
- A single infrastructure provider. The same model, served by someone else, at a different numeric precision or on a different runtime, may produce different figures. That is exactly this article's conclusion, applied to the article.
9. Falsifiable predictions
Each prediction carries an explicit horizon, a probability stated in calibrated language, and a public source where it can be checked. They are worded so they can be contradicted.
| Prediction | Horizon | Probability | How to check |
|---|---|---|---|
| F1. The price actually billed for output tokens on GLM-5.3-Flash differs from today's ($0.25, with the promotion applied) | 1 December 2026 | very likely (0.80) | The provider's public price list |
| F2. An identical rerun yields, for at least one model, a median latency differing by more than 20% | 1 November 2026 | likely (0.70) | Repeating the protocol in this article |
| F3. Neither maker publishes a full evaluation protocol — prompts, run count, automated criterion | 1 January 2027 | very likely (0.80) | The published model cards |
| F4. On a rerun, the ordering of the two models reverses on at least one of the five axes | 1 November 2026 | more likely than not (0.60) | Repeating the protocol in this article |
| F5. The parameter requesting "no deliberation" keeps producing deliberation in the answer body | 1 November 2026 | more likely than not (0.65) | A single call with that parameter |
| F6. On a rerun of the analysis-and-forecast test, at least one model again produces already-expired deadlines in most runs | 1 November 2026 | very likely (0.85) | Repeating the sixth test from the article |
10. How to reproduce it
The prompts are the ones described in point 2. Parameters: temperature 0.0 for classification and 0.2 elsewhere; output budgets of 300 tokens for classification, 700 for JSON, 1,200 for the word-limited answer and 1,500 for the maths; two runs for the first three tests, one for the fourth; eight runs at temperature 0.3, with a 900-token budget, for the analysis-with-forecast. Scoring: word counting for the length limit, direct parsing for JSON, searching for the correct result at the end of the answer for arithmetic, and for the forecasts comparing the deadline the model wrote against the date of the measurement. The list prices, the active discount and the effective price are those the provider displayed publicly on the day of measurement.
Sources: the infrastructure provider's public model and price list; the model cards published by Z.ai; the model cards published by DeepSeek; the paper documenting the limits of grading by a third model — position, verbosity and self-enhancement biases; Regulation (EU) 2024/1689 on artificial intelligence.
Disclaimer: This material is for informational and analytical purposes. It does not constitute investment advice, a technology procurement recommendation or a legal assessment. The measurements describe a one-hour window on a single day, on a single infrastructure provider, with a small number of runs, and do not generalise without independent verification. The probabilities are explicit estimates, worded so events can contradict them, not certainties. Prices change frequently.