DeepSeek V4.1 Pro vs Qwen3-Max: Cheaper per Token, a Quarter of the Window

This one resolves on a single column, and it is not price. Qwen3-Max costs $0.36 per million input tokens and $1.43 output, serves a 262K-token context with a 65K output ceiling. DeepSeek V4 Pro costs $0.66 and $1.98 off-peak, serves a 1M context with 384K of output. DeepSeek V4.1 Pro, the name in the title, has no card at all.

So Qwen is 45% cheaper on input and 28% cheaper on output at DeepSeek's off-peak rate — and in DeepSeek's peak window the gap becomes 73% and 130%, in Qwen's favour. Against that, DeepSeek offers a context four times larger and an output ceiling nearly six times larger.

The cards

• Qwen3-Max — $0.36 in, $1.43 out; 262K context; 65K maximum output.

• DeepSeek V4 Pro — $0.66 in, $1.98 out off-peak; $0.022 cached; $1.32 and $3.96 during two weekday windows; 1M context; 384K output; text-only; 6.22 seconds to first token; 500 concurrency.

• DeepSeek V4.1 Pro — named once, in September, and never priced.

When the cheaper model is the wrong model

Two hundred and sixty-two thousand tokens is roughly 200,000 words — a medium codebase, a long contract set, a multi-hour transcript. Sixty-five thousand output tokens is about 50,000 words. Both numbers are generous for chat and tight for generation.

If your workload lives inside those bounds, Qwen wins on every task that touches it, and by more than the sticker suggests once peak windows are counted. If it does not, Qwen does not discount the job — it refuses it, and you pay for the chunking pipeline instead.

The honest test is the 95th percentile of your input length, not the average. The average never binds; the tail is where the expensive surprise lives.

What the price difference pays for on the DeepSeek side

Three things, and only one of them is a number.

• The window — 1M against 262K, and 384K of output against 65K.

• The cache line — $0.022 per million on a hit, against $0.66 on a miss. Qwen3-Max's card as we read it publishes no equivalent, which means an agent that re-sends a stable prefix cannot price that behaviour in advance.

• A published licence — MIT for DeepSeek V4 Pro with weights available. Qwen3-Max's weights are not published in the same form.

Against that, Qwen wins latency by default: no comparable first-token figure is on its card, but nothing in its pricing suggests a trillion-parameter backbone. DeepSeek's 6.22 seconds is the slowest number in this comparison and it does not move.

Per task

One million input tokens and 100,000 output tokens. Qwen3-Max: $0.36 plus $0.143 — about 50 cents. DeepSeek off-peak: 86 cents. In a peak window: $1.72.

So the same task is 50 cents, 86 cents or $1.72 depending on model and hour — a 3.4× swing driven entirely by DeepSeek's schedule. For a workload that can be moved, that is a lever. For one that cannot, Qwen's flat card wins by default.

What to run

Qwen3-Max and DeepSeek V4 Pro are both on OrcaRouter at their vendors' published rates, on one OpenAI-compatible endpoint with automatic failover — which makes the test a routing rule rather than two integrations.

The rule: short-input, short-output volume to Qwen at 36 cents; anything in the long tail, or anything with a cacheable prefix, to DeepSeek in its off-peak window. And V4.1 Pro stays out of the table until DeepSeek publishes an identifier, a price column or a repository — none of which has moved as of 30 October.