A generated title card reading 'Qwen4-Max vs DeepSeek V4.1 Pro' with the subhead 'One has a rate card. The other has a sentence in a changelog.', over three cards: Qwen4-Max at 1M context and $1.50 / $4.50 per 1M; DeepSeek V4.1 Pro with no model ID and no price, no card; DeepSeek V4 Pro at $0.66 / $1.98 per 1M with AA Index 23. A footer reads 'Qwen4-Max figures per Alibaba and Artificial Analysis; DeepSeek V4.1 Pro is unreleased as of 16 November 2026.'
Guides & Insights

Qwen4-Max vs DeepSeek V4.1 Pro: One Has a Rate Card, the Other Has a Clause

Author

jinhao.song

Date Published

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

Qwen4-Max went metered on 16 November 2026: $1.50 per million input tokens and $4.50 per million output in Singapore, a 1,000,000-token context window, native image and video input, and an Artificial Analysis Intelligence Index of 43 — 38th of 81 models on that board. D​eepSeek V4.1 Pro has no model identifier, no price row, no context window, no benchmark and no weights, because D​eepSeek has never released it. The only public document that names it is the 10 September 2026 release note for DeepSeek V4.1 Flash, which said that from 14 September requests to deepseek-v4-pro would be routed to V4.1 Flash "until V4.1-Pro launches" — a plan D​eepSeek withdrew after developers objected to a production model being swapped underneath them. Two months later the name still has nothing behind it. So the comparison worth your afternoon is not Qwen4-Max against a model that does not exist; it is Qwen4-Max against the D​eepSeek flagship that does.

What shipped on 16 November, in the numbers a buyer uses

A​libaba's release describes Qwen4-Max as the first model on the Qwen4 architecture: a sparse mixture-of-experts with 2.4 trillion backbone parameters, 71 billion activated per token, plus 328 billion more in N-gram embedding parameters that are looked up rather than multiplied. Three Gated DeltaNet layers compress history into fixed-size state for each Q​wen Sparse Attention layer, and four residual branches with dynamic read and write gates hold the deep stack together. That is the vendor's account of why long-context cost falls; it is an account, not a measurement.

• Context — a 1,000,000-token window, with an input ceiling of 991,808 tokens in non-thinking mode and 983,616 in thinking mode. Artificial Analysis lists the model at 984k, which is the same window measured from the other end.

• Output — 131,072 tokens non-thinking; in thinking mode reasoning and answer share a single 262,144-token budget. reasoning_effort accepts low, high and xhigh, and omitting it gives you high.

• Price, Singapore — $1.50 in, $4.50 out per million. Cache reads at $0.075, cache writes at $1.875 for five minutes or $3.00 for an hour.

• Price, everywhere else — $1.413 in, $4.24 out across Frankfurt, Virginia, Tokyo, Hong Kong and Beijing. Same model, 5.8% cheaper per token than the Singapore card.

• Against its own predecessor — Qwen3.8-Max lists at $2.00 and $6.00 in Singapore, so the generation-over-generation cut is 25% on both sides of the meter.

• Modality and surface — text, image and video in, text out; O​penAI Chat Completions, O​penAI Responses, A​nthropic Messages and DashScope; batch inference in Beijing only.

Every one of those figures is A​libaba's own. The one independent number available today is the AA Index of 43, which puts Qwen4-Max above the board's median of 41 and below the frontier tier. Its cost to run that index once is $3.27, blended price $0.80 per million, and it emitted 127 million output tokens doing it — verbose, but not exceptionally so.

What the V4.1 Pro name actually consists of

Less than a paragraph. D​eepSeek's 10 September note introduced DeepSeek V4.1 Flash, a 552-billion-parameter mixture-of-experts on a new Causal Encoder-Decoder architecture that activates 8 billion parameters during prefill and 16 billion during decode, reads images, serves a 1M-token context and 384K of output, and shipped under MIT. In the same note D​eepSeek wrote that V4 Pro traffic would move to Flash at Flash rates "until V4.1-Pro launches." Developers read that as a retirement notice for a model they were running in production; D​eepSeek backed the routing plan out and kept serving deepseek-v4-pro at unchanged billing.

What has appeared since is nothing verifiable. D​eepSeek's pricing page today lists exactly two served models: deepseek-flash, mapping to DeepSeek-V4.1-Flash-1017, and deepseek-v4-pro, mapping to DeepSeek-V4-Pro-0813. There is no deepseek-v4.1-pro string in the API reference, no row on the price sheet, and the newest repository in the deepseek-ai organisation on Hugging Face is still V4.1 Flash. Trackers that publish a date for it are publishing an estimate; one widely shared countdown page lists 12 November 2026 as an "estimated release," and that date passed without an announcement.

So the honest reading of the name is this: it is a tier D​eepSeek has described in prose and has not built anything public on top of. The V4 generation is the only evidence for what the tier would look like — DeepSeek V4 Pro is a 1.6-trillion-parameter mixture-of-experts with 49 billion active, text-only, 1M context, 384K output, MIT licence — and a V4.1 Pro would presumably inherit the new architecture and native vision. Presumably. The vendor's own claim that Flash "comprehensively surpasses" V4 Pro on performance, cost and speed leaves the Pro tier's purpose in the V4.1 line genuinely unclear, which is a strange position for a flagship to be announced into.

A generated two-column scoreboard titled 'Qwen4-Max vs DeepSeek V4 Pro — the scoreboard', using the same six row labels on both sides. Qwen4-Max: context 1,000,000 tokens, output 262,144 thinking plus answer, price per 1M $1.50 / $4.50, modality text, image and video, AA Index 43, cost per task $3.27. DeepSeek V4 Pro: context 1,048,576 tokens, output 384,000 tokens, price per 1M $0.66 / $1.98, modality text only, AA Index 23, cost per task $1.77. Footer: 'Price and context vendor-reported; AA Index and cost per task per Artificial Analysis v5.0.1.'

The matchup you can run today: Qwen4-Max against DeepSeek V4 Pro

Strip the version number off the D​eepSeek side and there is a real contest, because DeepSeek V4 Pro 0813 is a callable, priced, open-weight flagship. Both sides below are the vendors' published figures; the index column is Artificial Analysis v5.0.1, which is the only row where the two numbers were produced by the same harness.

• Price per million — Qwen4-Max $1.50 in / $4.50 out (Singapore) vs DeepSeek V4 Pro $0.66 / $1.98 off-peak, doubling to $1.32 / $3.96 in the weekday peak windows of 01:00–04:00 and 06:00–10:00 UTC.

• Context and output — 1,000,000 tokens with a 262,144-token thinking-plus-answer budget vs 1,048,576 tokens with up to 384,000 output tokens.

• Modality — text, image and video in vs text only. This is the single largest functional difference between them and it is not close.

• Backbone — 2.4T total, 71B active, plus 328B in N-gram embeddings vs 1.6T total, 49B active. Parameter counts predict training cost, not token efficiency; treat them as architecture notes.

• Independent index — 43 vs 23 at D​eepSeek's Max effort setting, with 21 at High and 13 non-reasoning. Twenty points is a generation-shaped gap, not a settings-shaped one.

• Independent cost per index task — $3.27 vs $1.77. Note the direction: D​eepSeek wins on cost, by roughly 1.8×, on the evaluator's own token accounting.

• Verbosity — 127M output tokens across the index vs a figure D​eepSeek's page reports per task rather than in total; the comparison that matters is cost per completed task, not tokens per token.

• Weights — proprietary, API only vs MIT, downloadable in 48 safetensors shards.

• Cache economics — Qwen4-Max charges $1.875 per million to write a five-minute cache entry and $0.075 to read it, a 95% discount; D​eepSeek charges nothing explicit to write and $0.022 per million to read, a 97% discount. D​eepSeek's cache is both cheaper and automatic, which matters more than the headline rate on agent workloads.

Costing a long-context job on both meters

Take a 300,000-token codebase prefix, ten turns of work against it, 4,000 output tokens per turn, thinking on, and a prefix that is stable across turns. This is arithmetic on published list prices, not a measured run.

• Qwen4-Max, Singapore, uncached — 300,000 input tokens at $1.50 per million is $0.45 per turn; ten turns is $4.50. Output is 4,000 tokens at $4.50 per million, $0.018 per turn, $0.18 across ten. Total $4.68.

• Qwen4-Max with the cache doing its job — one write at $1.875 per million, $0.5625, then nine reads at $0.075 per million, $0.0225 each, $0.2025. Input falls to $0.765, output stays $0.18. Total $0.945, a 3.8× cut for one line of request configuration.

• DeepSeek V4 Pro, off-peak, uncached — 300,000 input tokens at $0.66 per million is $0.198 per turn, $1.98 across ten. Output at $1.98 per million is $0.00792 per turn, $0.0792 total. Total $2.06.

• DeepSeek V4 Pro with cache hits — the first turn pays $0.198; the nine later turns read at $0.022 per million, $0.0066 each. Input lands near $0.257, output $0.079. Total $0.336. In the peak windows, double it.

Two conclusions survive the arithmetic. First, at list price Qwen4-Max costs roughly 2.8× what DeepSeek V4 Pro costs on this shape of job, cached or not. Second, A​libaba's claim that Qwen4-Max uses 24% to 35% fewer tokens than Qwen3.8-Max does not close a 2.8× gap; even a generous 35% token saving leaves Qwen4-Max at about twice the cost per completed task. The vendor's own $4.43-to-$1.67 per-task comparison is a comparison against its own previous generation, not against D​eepSeek.

What would make V4.1 Pro a real opponent

Three published artefacts, and they are the same three that have signalled every D​eepSeek release. A changelog entry naming the model, since that is where V4.1 Flash and V4 Pro 0813 were both announced. A repository in the deepseek-ai organisation with V4.1-Pro in the name, which is where the weights would land if the tier follows the family's MIT convention. Or a third column on the pricing page, which today has exactly two models on it. Until one of those moves, any spec sheet you read for D​eepSeek V4.1 Pro — parameter count, context, price, score — is extrapolated from the V4 generation and should be read that way.

There is one more thing to watch that is not an artefact but a claim: D​eepSeek says Flash beats V4 Pro on performance, cost, speed and total task time. If that stays true at the V4.1 generation, a V4.1 Pro has to justify itself on something other than "bigger" — and the most likely answer is the one nobody has published, which is that it will be the multimodal, long-horizon tier rather than the cheap one.

Getting either one behind a single key

OrcaRouter does not serve Qwen4-Max. It is not in the catalogue of more than 200 models as of 16 November, and no adjacent listing makes it so; the honest route to Qwen4-Max today is A​libaba's own Model Studio API and several third-party platforms. What we do serve is the D​eepSeek side of this page, at the vendor's list price passed through with no markup: deepseek/deepseek-v4-pro-0813 at $0.66 in and $1.98 out per million, 1M context, 384K output, the O​penAI, Responses and Anthropic-compatible endpoints, with the off-peak and peak windows applying exactly as D​eepSeek publishes them.

That asymmetry is the practical point of this comparison. You can put DeepSeek V4 Pro on the key you already have and measure it this afternoon. You cannot do that with Qwen4-Max here yet, and pretending otherwise would be the kind of claim that costs a team a week. What a routing layer does buy you in the meantime is a cheap mixed state: DeepSeek V4 Pro carrying the volume at list price, a shadow slice of traffic sent to A​libaba directly for Qwen4-Max, and a single place where the two bills are read side by side. When a model does land in the catalogue, moving a percentage of traffic to it is a routing rule rather than an integration, and automatic failover means an unproven model can sit behind a proven one without becoming a single point of failure on its first bad afternoon.

A screenshot of the OrcaRouter model page for DeepSeek V4 Pro 0813 at the identifier deepseek/deepseek-v4-pro-0813, showing the $0.66 input and $1.98 output prices, the Tools, JSON and Reasoning capability tags, the attribution to DeepSeek, and the description of a 1M-token context window with up to 384K output tokens, thinking and non-thinking modes with low, high and max reasoning effort, and native support for the Responses API.

The decision in front of you this week

If you need the strongest model you can call today and cost is the constraint, DeepSeek V4 Pro is the answer at $0.66 and $1.98, and the twenty-point index gap to Qwen4-Max is the price of that discount — a real price, not a rounding error. If your workload has video or images on the input side, DeepSeek V4 Pro cannot do it at all, and Qwen4-Max's native vision plus a 1M window is worth the 2.8× even before you count the Work360 agent claims A​libaba has published and nobody has reproduced.

If what you actually wanted was D​eepSeek V4.1 Pro, the answer is that there is no such model to wait for, only a clause in a September changelog that D​eepSeek has not acted on. Build against what exists; keep the switch cheap; and check D​eepSeek's pricing page once a month, because that is the page that will change first.

A screenshot of Alibaba Cloud Model Studio's documentation page for qwen4-max (Latest Model), dated 11-16, showing the model description as the next-generation flagship multimodal model in the Qwen family covering text generation, image and video understanding and deep reasoning, its support for web search, tool calling and structured outputs, its OpenAI and Anthropic API compatibility, and its deployment regions including China (Beijing), Singapore and Germany (Frankfurt).

Compared in this article3

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