Kimi K3.1 Pro vs Qwen 4 Max: Only One Has a Rate Card

You can call Kimi K3.1 Pro right now, at $2.00 per million input tokens and $10.00 per million output, on a model ID that resolves. You cannot call Qwen 4 Max at all. Thirty-four days after Alibaba put it on a conference stage, there is no endpoint, no model card, no price, no context figure and no published benchmark — only a name, a tier position and the word "training".

That asymmetry is not a reason to skip the comparison. It is the comparison. Anyone deciding what to build on in late October 2026 is choosing between a model with a rate card and a promise, and the useful work is to say precisely what the promise contains, what it would have to beat, and which of those two things you can actually put in production this quarter.

What was announced, and what was not

Alibaba previewed the Qwen 4 family on September 22, 2026 at its Apsara conference in Hangzhou. Four tiers were named: Qwen 4 Max as the flagship, Qwen 4 Plus as the balanced mid-tier, Qwen 4 Flash for throughput, and a Qwen 4 27B for local deployment. The Qwen LLM lead described the line as in training and said only that it would arrive "very soon".

Everything else was left blank, and the blanks are the specification. No parameter count. No context window. No price. No API identifier. No benchmark score. No release date. The only number in the announcement is the 27 in the open-weight tier's name, which 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 with Qwen 4 belongs to the Qwen 4.5 and Qwen 5 generations. It is not a Qwen 4 Max specification, and no source inside Alibaba has attached it to this model. Any spec sheet you find with a parameter count, a 1M context and a SWE-bench score for Qwen 4 Max is a projection, not a reading.

There is a design preview worth reading, though. Qwen3.8-Flash-Next, open-sourced in late August 2026, is explicitly billed as an architectural look at the Qwen 4 generation: roughly 125B main parameters plus 51B of N-gram embedding parameters, about 6B active per step, a 262K native context extensible toward 1M, and a reported training cost around 90% below Qwen3.7-Plus. That tells you the direction — sparse, retrieval-augmented, cheap to train — without telling you the flagship's size or price.

The half of this comparison that exists

Kimi K3.1 Pro shipped on October 23, 2026. It is a post-trained successor to Kimi K3 rather than a new architecture: the same 2.8-trillion-parameter mixture-of-experts backbone with roughly 104B active parameters, the same 1,048,576-token window, the same native image and video input. Moonshot AI's claim is not a bigger model but a better-behaved one, trained with large-scale reinforcement learning across agentic and professional domains.

• Price — $2.00 input / $10.00 output per million tokens, with cached input at $0.20 and cache writes at $2.00 and $4.00 for five-minute and one-hour TTLs.

• Highspeed variant — $4.00 / $20.00 for two to three times the output speed, with the cache-hit rate unchanged.

• Context — 1,048,576 tokens in, up to 943,718 tokens out, text, image and video in, text out.

• Weights — promised by November 4, 2026, which would make it the highest-scoring open-weight model released to date.

The vendor's own benchmarks put it at 65.7% on Terminal-Bench 4.0, 92.7 on BenchCAD with tools, 64.4 on OfficeQA Pro v2 and 51.7 on SWE Marathon, with hallucinated claims falling from 23.4% to 8.3% when tools are available. Those are Moonshot's numbers on Moonshot's harnesses and should be read as claims, not measurements. What is closer to independent is Artificial Analysis's Intelligence Index, which places it at 55 — fourth overall, behind Claude Opus 5.5 at 58, and the highest reading any soon-to-be-open model has recorded against the frontier.

The matchup you can run this afternoon

If the real question is "Kimi or Qwen", the honest head-to-head is not against a slide. It is Kimi K3.1 Pro against Qwen3.8-Max, the Alibaba flagship that has been generally available since August 3, 2026 and is the model a Qwen customer is using today.

• Price — Kimi K3.1 Pro $2.00 / $10.00 against Qwen3.8-Max $2.00 / $6.00, with cached input at $0.20 and $0.25 respectively. Same input, 67% more on output for Kimi.

• Context — 1,048,576 tokens against 1,000,000. Effectively a tie on the commercial endpoints, and not a tie on the open weights, where the published Qwen3.8-2.4T-A95B checkpoint carries roughly 262K and text only.

• Open weights — Kimi's are due November 4; Qwen3.8-Max's have been downloadable since August 12, 2026, under a custom licence with revenue and user-count thresholds.

• Architecture — 2.8T total with ~104B active on Kimi; 2.4T total with ~95B active on Qwen3.8-Max. Close enough that neither number predicts a winner.

• Reasoning — Kimi reasons by default with selectable effort; Qwen3.8-Max exposes its own thinking controls through Alibaba's API.

There is no controlled head-to-head between these two on a shared harness, which is the single biggest gap in the public evidence and the thing to watch for. What exists instead is a set of indirect readings that disagree with each other. Artificial Analysis has scored Qwen3.8-Max anywhere from 40 to 56 depending on which revision of the index was in force, which is a useful reminder that a single number from that board is a snapshot, not a verdict. Kimi K3.1 Pro's 55 was taken under the current revision. Comparing the two as if the readings were contemporaneous would be exactly the kind of laundering this comparison is trying to avoid.

What Qwen 4 Max would have to clear

Three tests decide whether the wait is worth it, and none of them is a benchmark chart.

The first is delivery. Qwen's own recent record sets the bar: Qwen3.8 went from preview on July 19, 2026 to general availability on August 3 and open weights on August 12 — a full generation from preview to downloadable in under four weeks. If Qwen 4 Max follows that pattern, the gap between "announced" and "usable" is short. If it does not, the silence since September 22 is the signal.

The second is price. Qwen3.8-Max sits at $2.00 / $6.00 and Kimi K3.1 Pro at $2.00 / $10.00, so Alibaba already wins the output line against the model it is actually competing with. A Qwen 4 Max that costs more than its predecessor without a clear capability jump would be a harder sell than the launch coverage assumes, and a flat-price upgrade would be the most dangerous outcome for everyone else in the market.

The third is openness. The 27B tier is promised as open weights; the Max tier's licence is unconfirmed. Kimi's November 4 commitment is the thing Qwen 4 Max will be measured against, because a frontier-class flagship that is downloadable changes procurement decisions in a way an API-only model cannot.

What Qwen 4 Max does not have to clear is the suspicion that it is vapor. The Qwen 3.8 line shipped on schedule with weights, the architecture preview is public, and Alibaba has the training capacity. The reasonable prior is that Qwen 4 Max arrives and is competitive. The unreasonable position — and the one a lot of SEO content is taking — is to publish its spec sheet before Alibaba has.

The cost of waiting, by reader

If you are starting a new build this month and your workload is long-document analysis, agentic coding or anything that needs a million-token window today, Kimi K3.1 Pro is the defensible choice. It is callable, it is priced, and its weights are eleven days away at the time of writing. Waiting for a model with no date is not a strategy; it is a schedule risk with no upside you can quantify.

If you are already on the Alibaba stack — Qwen3.8-Max in production, tooling built around its API, a procurement relationship — the answer is different. Qwen3.8-Max is cheaper on output, its weights are already public, and the migration to Qwen 4 Max will be a model-string change rather than a re-platforming. The rational move is to keep building on what exists and treat the launch as an upgrade, not a decision point.

If you are choosing a vendor rather than a model, the asymmetry matters less than it looks. Both labs ship on roughly quarterly cadence, both publish weights, and both cut prices when pressured. The lock-in risk is in the integration, not in the model name — which is the argument for putting a routing layer between your application and either lab.

Hedging instead of guessing

OrcaRouter carries kimi/kimi-k3.1-pro and qwen/qwen3.8-max on one key at provider list price with no markup, so the prices quoted above are the prices billed rather than a reseller's version of them. Qwen 4 Max is not in the catalogue, and will not be until Alibaba releases it — there is nothing to route to.

What the routing layer buys you in the meantime is a way to write the decision down instead of arguing about it. A policy can send long-context document work to Kimi K3.1 Pro, high-volume generation to Qwen3.8-Max, and fail over between them automatically when a provider rate-limits or degrades. When Qwen 4 Max does ship, the change is a model entry in that policy rather than a new integration, and because pricing is passed through at list, a launch-day price cut is live on the same terms rather than after a repricing cycle.

That is a modest claim, and it is the only one worth making here: the honest hedge against an unannounced model is not a prediction, it is an architecture that does not care which of the two wins.

Three dates to watch

November 4, 2026 — Kimi K3.1 Pro's weights. This is the first date on which the model can be evaluated without an API budget, and the first on which the open-weight claim becomes checkable rather than promotional.

Qwen 4 Max's launch — no date has been given. Watch for a model card rather than a keynote: parameters, context, price and an API identifier in the same document is the signal that the model is real. Until that exists, treat every spec sheet as speculation, including the ones that look authoritative.

The first shared-harness comparison — the moment someone runs Kimi K3.1 Pro, Qwen3.8-Max and Qwen 4 Max through the same scaffold on the same tasks. Everything published so far is either a vendor number or an indirect reading, and the difference between the two is the entire reason this article has a caveat section.

Until those three dates land, the accurate one-line answer is the one the evidence supports: Kimi K3.1 Pro is the model you can run, Qwen 4 Max is the model you can wait for, and Qwen3.8-Max is the model most Qwen customers will actually be comparing against this quarter.