A generated 16:9 title card on a dark charcoal background with faint horizontal rules, headlined 'Qwen4-Max vs GLM-5.4' and subtitled 'One shipped. One missed its window.' A vertical timeline of dated dots runs down the right side at 2026-02-12, 2026-04-07, 2026-06-16, 2026-08-18 and 2026-08-26, ending in an empty dashed segment labelled GLM-5.4. A teal badge at bottom left reads 'window expired 27 Oct 2026'. OrcaRouter logo bottom-right.
Guides & Insights

Qwen4-Max vs GLM-5.4: One Shipped, One Missed Its Own Window

Author

jinhao.song

Date Published

Latest models · 20View all models →
Benchmarks: Artificial Analysis · updated daily
Back to all posts

There is a version of this comparison that would be useful and a version that would be fiction. The useful one starts with a date: Z.ai's own release log still ends at GLM-5.3-Flash on 26 August 2026, its documentation index enumerates GLM-5.2, GLM-5.3 and GLM-5.3-Flash and nothing above them, and the GLM-5.4 documentation path returns Page Not Found as it did in October. Qwen4-Max, by contrast, has been on metered sale since 16 November 2026, with a published rate card and an independent score. So the matchup is not two flagships. It is a flagship and a projected date that has now expired.

What makes that more than a debunking is the arithmetic behind the projection. This blog put a 54-to-70-day cadence on the G​LM line in October and published the window it implied. That window has closed. Working out by how much, and what the closure tells you about how Z.ai is sequencing releases, is genuinely more informative than another speculative spec sheet — and it leaves a comparison that can actually be run: Qwen4-Max against GLM-5.3, the G​LM that ships, on the same independent index, in the same week.

Where the name stands on 18 November

The audit is short and it is repeatable. Four surfaces, all checked today, all negative.

• The release log — newest entry GLM-5.3-Flash, dated 2026-08-26; then GLM-5.3 on 2026-08-18, GLM-5.2 on 2026-06-16, GLM-5.1 on 2026-04-07, GLM-5 on 2026-02-12

• The documentation index — GLM-5.3-Flash/FlashX, GLM-5.3, GLM-5.2. No 5.4 entry at any effort level

• The documentation path — docs.z.ai/guides/llm/glm-5.4 returns 404, exactly as it did on 13 October

• Hugging Face — a repository search for "glm-5.4" returns nothing under the lab's own organisation, and the newest text checkpoints there are the four GLM-5.3 releases created on 25 August

The vendor's model-api page gives the same answer from the commercial side: GLM-5.3, GLM-5.3-Flash, GLM-5 and GLM-5V-Turbo, with no row between 5.3 and 6. Nine surfaces were checked for the standing page on this name in October; the same four that matter for shipping status are still negative five weeks later. That consistency is the finding. A name that has been absent from every first-party surface for eight weeks is not a launch in flight; it is a name the line has passed over, at least for now.

A generated two-column scoreboard titled 'Qwen4-Max vs GLM-5.3' with six rows: input per million $1.50 against $1.40; output per million $4.50 against $4.40; index score 43 against 32; cost per task $3.27 against $5.01; wall-clock per task 32.1 min against 22.0 min; weights none against downloadable. Footer reads 'Artificial Analysis Intelligence Index v5.0.1; prices per vendor list'. OrcaRouter logo bottom-right.

The window, and how far past it we are

The cadence argument was never a leak. It was arithmetic on Z.ai's own release dates: GLM-5 to GLM-5.1 took 54 days, GLM-5.1 to GLM-5.2 took 70, GLM-5.2 to GLM-5.3 took 63. Apply the same span to 18 August and the next text flagship was due somewhere between 11 October and 27 October 2026. If you date the family from the Flash tier's 26 August entry instead, the outer edge moves to 4 November.

It is 18 November. Both readings have expired, the later one by two weeks. That is the cleanest statement available about GLM-5.4 today, and it is stronger than "we could not find it", because it is falsifiable: a single dated entry on the vendor's release log would overturn it, and until one appears the projection has been tested and has failed.

There are two honest readings of that failure, and they point in opposite directions. The first is that Z.ai has slowed down — that a 753-billion-parameter flagship is now expensive enough to iterate on a longer cycle, and the cadence of 2026's first half was a sprint that has ended. The second is that the next release is not a 5.4 at all: that the line skips a number, or that the effort has gone into the multimodal Flash tier, which is where the only post-August activity on the release log actually sits. Nothing published by the lab distinguishes between those two, and a page that guessed would be doing exactly what a scoreboard with invented columns does.

What you can actually compare: Qwen4-Max against GLM-5.3

GLM-5.3 is not a consolation prize. It is a 753-billion-parameter mixture-of-experts activating roughly 40 billion parameters per token, text-only, with a 1M-token context window and a 128K output ceiling, and it has something Qwen4-Max's launch coverage mostly lacked: a measured row on the same index, taken by the same harness, in the same index revision.

• Input price — Qwen4-Max $1.50 per million tokens in Singapore, $1.413 in its other five regions, versus GLM-5.3 $1.40 on Z.ai's list

• Output price — Qwen4-Max $4.50 versus GLM-5.3 $4.40

• Independent score — Qwen4-Max 43 on Artificial Analysis Intelligence Index v5.0.1 versus GLM-5.3 at 32 at max effort and 23 at low effort

• Cost per index task — Qwen4-Max $3.27 versus GLM-5.3 $5.01. This is the row that matters, and it points the opposite way from the score

• Wall-clock per task — Qwen4-Max 32.1 minutes versus GLM-5.3 22.0 minutes at max effort, 12.2 at low

• Token appetite — Qwen4-Max 127 million output tokens per index run versus GLM-5.3 135 million at max effort, 63.3 million at low

• Throughput — Qwen4-Max 55 output tokens per second versus GLM-5.3 90

• Modality — Qwen4-Max takes text, image and video and returns text; GLM-5.3 is documented as text-only, with visual work pushed to the Flash tier

• Max output — Qwen4-Max 131,072 tokens, or 262,144 shared between thinking and answer; GLM-5.3 128,000

• Weights — Qwen4-Max publishes none and marks fine-tuning and deployment unsupported; GLM-5.3's weights are downloadable under a custom Z.ai licence

• Effort control — Qwen4-Max exposes selectable levels; GLM-5.3 always reasons, with low, high and max

The price rows are almost a tie — a dime on input, a dime on output — and that is the least interesting thing in the list. The interesting thing is that the two models arrive at nearly identical list prices with different shapes underneath. Qwen4-Max costs more per task despite a cheaper meter because it takes longer and thinks longer; GLM-5.3 costs more per task despite a cheaper meter of its own because at max effort it burns through tokens at a rate the lower-effort setting does not. Neither vendor's price card tells you that. Only the cost-per-task row does, and it is the row both launch posts leave out.

The score gap is real but narrower than 43 versus 32 suggests, for a reason worth stating plainly: those figures come from different effort settings on the same index, and this blog has watched revision changes move a model's number by double digits. Treat 43 and 32 as a directional gap taken on one instrument on one day, not as a measured 11-point capability difference. What survives that caution is the cheaper claim: Qwen4-Max answers harder tasks than GLM-5.3 at max effort, and does it for less money per task, while being the slower model per token and the more expensive one per million.

A screenshot of Z.ai's release notes page headed 'New Released' with a right-hand date rail ending at 2026-08-26 and, below it, 2026-08-18, 2026-06-16 and 2026-04-07, and body entries describing native visual capabilities, a hybrid attention architecture with 320B total and 18B activated parameters, and GLM-5.3's coding gains on Z.ai Code Bench.

The thing a GLM 5.4 searcher actually needs

People type "GLM-5.4" for one of two reasons, and the two want opposite answers.

The first is shopping: they want to know whether to migrate off GLM-5.3, or whether to pick G​LM over Q​wen for a build starting this month. The answer is that there is nothing to migrate to, and the decision they can make today is between a text-only model with downloadable weights at $1.40/$4.40 and a multimodal rental at $1.50/$4.50 that scores higher and costs less per finished task. That is a real choice, and it does not require a 5.4 to make it.

The second is tracking: they want a date. The honest answer is that the cadence-based window for this name expired on 27 October, or 4 November on the most generous reading, and that the projection's failure is itself the update. The next signal to watch is not a leak but the release log: an entry above GLM-5.3-Flash would settle the question in one line, and it has not appeared in the twelve weeks since the Flash tier shipped.

Running either of them without a second contract

Here the two sides of this comparison genuinely differ, and the difference is about access rather than capability. OrcaRouter routes z-ai/glm-5.3 today, alongside the Q​wen 3.8 family, on one OpenAI-compatible endpoint at 0% markup — nothing added on top of what the serving provider charges, which on our own model page is $1.26 in and $3.96 out per million, at or under Z.ai's $1.40 and $4.40 list. What OrcaRouter does not route is Qwen4-Max: the model is not in the catalogue, and this page will not imply otherwise. Qwen4-Max is reachable through A​libaba's own six regions and through third-party platforms that resell it; if you want it on the same key as the G​LM, that is a question of waiting for the catalogue to carry it, not of configuring anything today.

A screenshot of the OrcaRouter model page for GLM 5.3, showing the identifier z-ai/glm-5.3, the byline Z.ai with the date 2026-08-18, a description of the model as a text-in text-out flagship for repo-scale coding, pricing of $1.26 input and $3.96 output per million tokens, a 1M-token context and a 128K maximum output.

Where routing does pay off immediately is the comparison itself. The cost-per-task numbers above are averages over a fixed benchmark mix, and your mix is not that mix. Running the same prompts through GLM-5.3 and a Q​wen 3.8 tier on one endpoint is how you find out which side wins on your traffic, with automatic failover covering you if one provider rate-limits mid-run — and it is a morning's work rather than two procurement conversations.

What would make this a real matchup

One document. A model card carrying parameters, context length, price and an API identifier together, dated on Z.ai's own release log. That is the standard this blog has applied all year, and it is the only thing that converts GLM-5.4 from a version number into a model. A second qualification, weaker but useful: a GLM-5.3 row measured at low effort against Qwen4-Max at its own low setting, which would let the cost comparison be made effort-for-effort rather than across the two vendors' defaults.

Until then the correct summary is uncomfortable for both sides of the search. Qwen4-Max is real, priced, measured, and slightly better than its reputation on cost — 43 on the index at $3.27 a task is a competent flagship, not a dominant one. GLM-5.4 is a number in a sequence whose expected date has passed without an entry on the log, and the G​LM you can run today is one digit lower, measurable, downloadable, a dime cheaper on the meter and half again as expensive per finished task at full effort. Anyone telling you otherwise is selling a spec sheet they cannot source.

Compared in this article1

Detected from this article · Benchmarks: Artificial Analysis · updated daily