Gemini 4 Carbon vs Qwen 4 Max: What Waiting Costs, Priced Three Ways

A​libaba has been showing Q​wen 4 Max on a conference stage for seven weeks and has not shipped it. The model went on stage at the Apsara conference in Hangzhou on 22 September 2026 as the flagship of a four-tier family, was described by Q​wen's LLM lead as training on a new-generation architecture, and has since produced no model card, no API identifier, no context window, no price and no benchmark score. G​emini 4 Carbon has been on general sale since 10 November 2026 at $2 and $10 per million tokens. So the comparison a reader actually needs is not two models. It is G​emini 4 Carbon against Qwen3.8-Max — the Q​wen flagship you can call this afternoon — plus an honest account of what waiting for Q​wen 4 Max would have to buy you. And because one side of this comparison reprices on 1 December, the cost of waiting is not a metaphor. It is a number.

What was announced, and what the announcement is worth

Four tiers were named at Apsara, and the roster is the entire published content of the announcement:

• Q​wen 4 Max — the flagship tier, positioned against the top model from every other frontier lab.

• Q​wen 4 Flash — the high-throughput tier, for latency-sensitive and high-volume work rather than peak reasoning.

• Q​wen 4 Plus — the balanced middle tier, historically the one with the widest multimodal spread.

• Q​wen 4 27B — the open-weight local tier, the one that gets downloaded, quantised and fine-tuned.

Everything else was left blank, and the blanks are the specification. No parameter count. No context window. No price. No API identifier. No published score. No date beyond "soon". The only concrete number in the announcement is the 27 in the open-weight tier's name, and it says something about that model's size and nothing about the flagship.

One figure needs retiring because it is already being repeated as fact. The 5-to-10-trillion-parameter roadmap that circulates alongside Q​wen 4 belongs to the Q​wen 4.5 and Q​wen 5 generations — it is a forward-looking statement about later models, not a Q​wen 4 Max specification. A​libaba has attached no parameter count to this model. Any spec sheet you find with a parameter count, a 1M context and a SWE-bench score for Q​wen 4 Max is a projection.

There is one piece of real evidence, and it is worth reading carefully. Qwen3.8-Flash-Next, open-sourced in late August 2026, is labelled on its own model card as a preview of the Q​wen 4 architecture. That makes it the only concrete thing anyone outside A​libaba knows about how the family is built: a sparse model with 125 billion main parameters plus 51 billion parameters of N-gram embeddings, activating roughly 6 billion per inference step, a 262,000-token native context that the card says extends toward 1 million, and a claim of roughly 90 percent lower training cost than Qwen3.7-Plus. Three techniques are named — QSA, a Qwen-specific sparse attention scheme; gated residual connections; and N-gram embeddings as a separate lookup store. It is a design direction, not a flagship.

The Q​wen you can call, and how it stands against Carbon

Qwen3.8-Max has been generally available since 2 September 2026 and is the top text model in A​libaba's own Model Studio documentation, which last updated on 2 November 2026 and does not mention a 4-series Max at all. Set it against Carbon on the dimensions that decide a decision.

• List price — G​emini 4 Carbon $2 in and $10 out per million tokens through 30 November 2026, then $4 and $20. Qwen3.8-Max $2.00 and $6.00, with cache reads at $0.25.

• What we charge — qwen/qwen3.8-max is in the catalogue at $2.00 and $6.00, and qwen/qwen3.8-max-0902 is the dated snapshot. Carbon is not in the catalogue and is reached through G​oogle's paid API.

• Context — 1,000,000 tokens for Carbon on G​oogle's docs; 984,000 for Qwen3.8-Max on Artificial Analysis's listing. Both vendors round to a million in marketing; the measured figures are what they are.

• Output ceiling — Carbon 65K with long-decode generation up to 1M; Qwen3.8-Max's maximum completion is not published on our catalogue entry and A​libaba's page does not state it in the summary. One-sided row, and worth checking against the vendor before you rely on it.

• Input modality — Carbon text, image, audio, video and PDF; Qwen3.8-Max text, image and video on the catalogue's modality string, versus text-only on the Artificial Analysis listing. The discrepancy is between a vendor-side capability string and an evaluator's summary; where they disagree, the vendor's documentation is the one to test against.

• Parameters — Qwen3.8-Max is a 2.4-trillion-parameter mixture of experts with 95 billion active per token, confirmed on the official model card. Carbon: undisclosed.

• Independent score — G​emini 4 Carbon 60 on Artificial Analysis Intelligence Index v5.0.1 at high effort, fourth of 74. Qwen3.8-Max 34, 46th of 74.

• Cost per index task — $7.76 for Carbon against $10.02 for Qwen3.8-Max, which is worth pausing on: the model that costs less per token costs more per completed task, because it is both slower and more verbose. Qwen3.8-Max consumed 150 million output tokens on the index against Carbon's 205 million, but it did so at 50 tokens per second rather than 79, and 57.3 minutes per task against Carbon's 26.

• Time to first answer token — 67 seconds for Carbon, 114 for Qwen3.8-Max.

That per-task inversion is the most useful number in this section and the one least likely to appear in a comparison table. On the headline card, Qwen3.8-Max is 40% cheaper on output than post-December Carbon. On the only measure that accounts for how much a model writes and how long it takes, it is 29% more expensive. Neither figure is a lie; they answer different questions, and the per-task one is closer to what a bill looks like.

Pricing the wait, which is the part nobody does

Waiting for Q​wen 4 Max is a decision with a cost, and unlike most wait-or-buy questions this one can be computed, because Carbon's price change is published.

• Scenario one: you build on Carbon now. November costs $2/$10, December costs $4/$20. A pilot that costs $200 in November costs $400 in December for identical traffic.

• Scenario two: you wait for Q​wen 4 Max. You spend November and December on whatever you have, and you pay no Carbon premium — but you also ship nothing, and every week of delay is a week of the same engineering work not being tested in production.

• Scenario three: you build on Qwen3.8-Max now and migrate if 4 Max lands. This is the only option with no cliff: Qwen3.8-Max has held $2.00/$6.00 since launch, there is a dated snapshot at qwen/qwen3.8-max-0902, and a successor arriving does not change the price of the model you are already calling.

The third option is usually correct for a reason that has nothing to do with benchmarks: it is the only one where the migration trigger is a new model appearing rather than a price changing underneath you. A team that standardises on a model whose price doubles in three weeks has to re-run its cost model in three weeks. A team on a stable card does not.

There is a symmetric cost to waiting that a table cannot capture. A​libaba's own family history says a Q​wen 4 Max launch will be followed quickly by tier siblings — Flash, Plus, and the 27B open release — and the 27B is the one with a predictable delivery record, because Q​wen's open tiers have consistently landed with MLX conversions, NVFP4 quantisations and community fine-tunes within days of release. If what you actually want is a Q​wen you can self-host, the tier to watch is not the flagship, and it is also not the one with no date.

Carbon's two real advantages, and the one that is situational

The first is the score, and it is not close: 60 against 34, measured by the same evaluator on the same revision of the same index, at each model's highest effort setting. Whatever a Q​wen 4 Max eventually posts, the bar it has to clear is that 60, not the 34 of its predecessor.

The second is the long-context economics. Carbon charges the same per-token rate at 2,000 tokens and at 900,000, with no length tier. That is a genuinely unusual property among 1M-context models, several of which step up a price band past a threshold — and it is the property that matters most for the workload Carbon is sold for, which is long-horizon work over large codebases and document sets.

The situational advantage is cyber and scientific reasoning, where G​oogle's launch material makes claims no one else is making: SEC-bench Pro 88.2%, ExploitGym 54.8%, CWE-bench 71.0%, Project Fortify at 100% scan recall and 95% triage precision, and a 40-fold average increase in homologs retrieved across seven low-homology CASP17 protein families. All vendor-reported, none reproduced, on a model one day into public availability. They are also the claims that would justify the price if they hold.

Against that, Qwen3.8-Max's case is narrower and more durable: a published 2.4-trillion-parameter card, a stable price, a dated snapshot for reproducibility, and a cost per token that lets you route volume work to it without a second thought. It is not a Carbon competitor on capability. It is a Carbon complement on economics, and for most pipelines that is the role that gets called more often.

The tier that may matter more than the flagship

Q​wen 4 27B is the least discussed name in the announcement and the one with the best delivery record. It is the only tier in the family that has historically shipped under open weights, and the Q​wen 3.8 generation showed what that unlocks: within days of the 27B weights landing, the community had MLX conversions, NVFP4 quantisations and uncensored fine-tunes. An announced 27B is a promise of a downstream ecosystem, and it is the part of the roadmap a self-hosting team should be tracking — not the flagship whose only published attribute is a tier position.

The flagship matters for a different reason. A​libaba positions Q​wen 4 Max against the top model from every other frontier lab, which means the comparison that will be run on launch day is Q​wen 4 Max against G​emini 4 Carbon at whatever Carbon costs in that month. If Carbon is at $4/$20 by then and Q​wen 4 Max lands near Qwen3.8-Max's $2/$6, the price argument flips hard and the capability argument is the only one left. That is a real possibility and it is the reason not to standardise exclusively on either.

What to do this month

Three moves, in order of how much they cost to get wrong.

Price every pilot on the December Carbon rate. Anything decided on November numbers will look twice as good as it is, and the correction arrives on a fixed date rather than an uncertain one. If a workload only makes sense at $2/$10, it does not make sense.

Keep a Qwen3.8-Max route warm rather than debating it. It is 40% cheaper per output token than December Carbon, it has a dated snapshot so the behaviour will not drift, and having it configured means the decision about Q​wen 4 Max becomes a swap rather than a project. Both models are on one key here — qwen/qwen3.8-max at A​libaba's list price with 0% markup, alongside the rest of a 216-model catalogue on the same OpenAI-compatible surface — and Carbon, again, is reached through G​oogle's paid API and not through us.

Do not build a capacity plan around Q​wen 4 Max. Nothing in the announcement carries a date, and the two firm facts about it are a conference stage and the word "training". When a model card appears with a price and a context window, the comparison can be re-run in an afternoon. Until then, a roadmap item is not a fallback, and a fallback you cannot call is not one either.