Gemini 4 Carbon vs GLM-5.5: One Side of This Matchup Has No Rate Card

There is a price for G​emini 4 Carbon and there is no price for GLM-5.5, because GLM-5.5 does not exist. Z.ai's release log runs GLM-5.3-Flash on 26 August 2026 and GLM-5.3 on 18 August, and stops. Its pricing page tops out at the same two models. There is no model card, no API identifier, no weight repository and no licence file anywhere in the G​LM line carrying the number 5.5 — and the company's own 31 August earnings call named GLM-6.0, not 5.4 or 5.5, as the next-generation base. So the useful comparison is not Carbon against a rumour. It is G​emini 4 Carbon, callable on the paid G​emini API since 10 November 2026, against GLM-5.3, the model Z.ai actually sells — plus an honest account of what a GLM-5.5 would have to beat to matter.

What exists, checked on 11 November 2026

Claims that something does not exist decay, so here is the method rather than only the conclusion. Every line is repeatable in an afternoon.

• Z.ai's release notes at docs.z.ai/release-notes/new-released — newest entry GLM-5.3-Flash, 26 August 2026, above GLM-5.3 on 18 August, GLM-5.2 on 16 June, GLM-5.1 on 7 April and GLM-5 on 12 February. Nothing numbered 5.4 or 5.5.

• Z.ai's pricing page — GLM-5.3 and GLM-5.3-Flash lead the list at $1.40/$4.40 and $0.15/$0.50 per million tokens. There is no GLM-5.5 row to price.

• Artificial Analysis — a live page for GLM-5.3, measured. No GLM-5.5 entry.

• Our own catalogue — 216 models, including z-ai/glm-5.3, z-ai/glm-5.3-flash and z-ai/glm-5.3-flash-free. No GLM-5.5, because there is nothing to serve.

That is four independent surfaces and they agree. The reason the name keeps appearing in search boxes is a separate question, and the short answer is that it was forecast: prediction markets and analyst notes expected a 5.5 through August and September, both of which passed with Z.ai shipping 5.3 and its Flash sibling instead. A number attached to an expectation is not a specification, and every table you find with a parameter count and a benchmark row for GLM-5.5 was generated from a slide rather than a model.

The comparison that can actually be run

GLM-5.3 is a serious model and it is the right stand-in, because it is what a reader typing "GLM 5.5" would end up calling today. Put the two against the same dimensions and the shape of the decision is clear.

• List price — G​emini 4 Carbon $2 in and $10 out per million tokens through 30 November 2026, then $4 and $20. GLM-5.3 $1.40 and $4.40, unchanged, with cache reads at 19% of input.

• What we charge — Carbon is not in our catalogue and is reached through G​oogle's paid API. z-ai/glm-5.3 is listed at $1.26 and $3.96 per million tokens, below Z.ai's own card, with z-ai/glm-5.3-flash at $0.15 and $0.50 for the throughput tier.

• Context and output — 1,000,000 tokens of context each. Carbon's documented output shape is 65K with long-decode generation up to 1M; GLM-5.3 caps output at 128,000.

• Input modality — Carbon takes text, image, audio, video and PDF. GLM-5.3 takes text and returns text. This is the largest single gap between them and it is not close.

• Parameters — GLM-5.3 is a 753-billion-parameter mixture of experts with 40 billion active per token, published on its model card. G​oogle has disclosed nothing about Carbon's size or architecture, and no technical report accompanied either the 28 October announcement or the 10 November opening.

• Thinking controls — GLM-5.3's reasoning cannot be switched off; thinking.type accepts only enabled, with effort at low, high or max and max as the default. Carbon's dial offers high, medium and low, and its index score moves with it: 60, 58 and 54.

• Independent score — G​emini 4 Carbon 60 on Artificial Analysis Intelligence Index v5.0.1 at high effort, fourth of 74. GLM-5.3 32, 43rd.

• Cost per index task — $7.76 for Carbon, $5.01 for GLM-5.3. That is the one axis where the cheaper model is also the cheaper task, and it is not enough to change the ordering.

• Speed — 79 output tokens per second for Carbon, 66 for GLM-5.3, with 67 seconds to first token against 58. Both are batch models.

The 28-point index gap is the number that ends the argument about raw capability, and it should be read with its caveats attached. It is one evaluator, one harness, one revision of one index, and GLM-5.3's 32 was measured at its max effort setting, which is its default. What it is not is a vendor claim: neither figure came from G​oogle or Z.ai, and both models were measured against their own list prices rather than a subsidised rate.

Where GLM-5.3 wins, and it is not nowhere

Three places, all of them about cost structure rather than capability.

The first is the price floor. At $1.26 and $3.96 through us, GLM-5.3 costs a third of what Carbon will cost after 1 December, and roughly two-thirds of Carbon's November introductory rate. For classification, extraction, routing, summarisation and the long tail of agent steps that do not need frontier reasoning, that difference compounds across a pipeline where the expensive model is called once and the cheap one is called forty times.

The second is that the price is stable. Carbon doubles on 1 December. GLM-5.3 has held $1.40/$4.40 since August, and a budget built on it does not need a footnote. In a quarter where one side of the comparison has moved its access terms twice and its price once, predictability is a feature you can put in a spreadsheet.

The third is that the weights exist. GLM-5.3 is downloadable under the GLM-5.3 licence. Carbon is a paid API with no weights, no published parameter count and no architecture. For a team whose compliance posture requires a self-hostable path, that is not a preference — it removes Carbon from the shortlist before any benchmark is consulted.

None of that closes a 28-point gap on a reasoning index. It does mean the honest recommendation is architectural rather than a winner: Carbon for the hard call, GLM-5.3 for the volume around it, and a routing layer between them so the split is a policy rather than a rewrite.

What a real GLM-5.5 would have to beat

This is the part worth keeping, because the name is going to become a model eventually and the bar is already set.

• The price — $1.40/$4.40 is the incumbent line. A 5.5 that arrives above it is asking existing G​LM users to pay more for a version bump, which Z.ai has not done across five generations.

• The score — 32 on the current index is the number to beat, and the meaningful target is not 33. It is the 60 that Carbon posts, because that is what a buyer comparing the two flagships is actually weighing.

• The modality — text-only is now the constraint that shows. Carbon reads audio, video and PDF; GLM-5.3 reads text. A 5.5 that stays text-only inherits the same disqualification on document-heavy work, whatever it scores.

• The evidence — a callable API identifier, published pricing with cache rates, an independent leaderboard entry, and either downloadable weights or a stated reason there are none. GLM-5.3 passes all four. A successor that launches with a conference slide and a parameter estimate does not, and the last two G​LM launches both shipped with the full set on day one, which is the standard to hold them to.

The parameter count that circulates with the rumour is worth one line, because it has drifted from one trillion to 1.6 trillion to 2.5–3.5 trillion across the year. A number that moves without a model attached is being estimated, not measured, and when the real card appears it will state its own figure.

The one line to take away

If you came here to choose between G​emini 4 Carbon and GLM-5.5, there is nothing to choose: one is a generally available model with a published rate card and an independent score, and the other is a version number that has never appeared on its own vendor's release log. Compare Carbon against GLM-5.3 and the decision is capability against cost, with Carbon ahead by 28 index points and behind by roughly two-thirds on price. Both models are cheap enough per token that the split is usually worth running rather than deciding — and running it on one key, with the vendor list price passed through at 0% markup so a Z.ai or G​oogle rate change lands the same day, is the version of this comparison that does not need rewriting when 5.5 finally ships.

When it does ship, the four-part test above is the one to apply. Until then, anyone quoting you a GLM-5.5 benchmark is quoting an estimate, and anyone quoting you a GLM-5.5 price is making one up.