Qwen4-Max Preview vs GPT-6.1 Sol: A Countdown Against a Rate Card

Put Qwen4-Max Preview and GPT-6.1 Sol next to each other on a spec sheet and they look like siblings: both claim roughly a million tokens of context, both take images, both reason before they answer, both speak the OpenAI wire format. Then look at the two commercial arrangements and the resemblance collapses. GPT-6.1 Sol was released on 29 September 2026 at $2.00 per million input tokens and $10.00 per million output, with cached input at $0.10 per million, and it is sold by the token to anyone with an account. Qwen4-Max Preview opened on 11 November 2026 as a Token Plan–only preview that runs until 24:00 on 25 November, has no published per-token price anywhere, and is retired when the window closes. One is a purchase. The other is a fortnight-long loan. That difference, not the benchmark column, is what should decide which one you open first.

Two machines, described the same way

GPT-6.1 Sol is a refresh of OpenAI's mid-tier reasoning model — the tier below GPT-6 Astra — and OpenAI's own framing is that it "nearly matches" Astra on agentic coding, computer use and professional work at a fifth of Astra's token prices. OpenAI's published model page puts the context window at 1,050,000 tokens with a 128,000-token maximum output and an April 30, 2026 knowledge cutoff. Those are vendor figures.

Qwen4-Max Preview is Alibaba's first callable Qwen 4 tier. It carries a 1,000,000-token context, a 393,216-token output ceiling shared between thinking and the final answer, text, image and video input, and function calling, web search and built-in tools. Those are vendor figures too, taken from the Model Studio documentation page rather than from coverage of it.

Read side by side, the two spec blocks differ by 50,000 tokens of context — noise on a number that size — and by a much larger margin on output headroom. 393,216 against 128,000 is roughly three times the room for a single response, and with thinking counted against that budget in both models, the difference is not academic. If your work is "read a long corpus, produce a long structured artifact", the Qwen preview has more rope. If your work is "answer well inside a normal completion", neither ceiling is your constraint.

The part that is not a price comparison

This is where the two pages stop being comparable at all, and it is worth being blunt about why rather than dressing it up as a tie.

GPT-6.1 Sol has a rate card you can do arithmetic on. Input at $2.00 per million, output at $10.00, cached input at $0.10 — one-twentieth of its own uncached input rate and a tenth of GPT-6 Astra's $1.00 cached rate. Agent loops that resend a stable prefix thousands of times live almost entirely on that last line, which is the quiet reason the 6.1 refresh matters to anyone running agents rather than chat. The card also has thresholds that change the arithmetic: cross 272,000 input tokens and the whole request reprices at twice the input and cache rates and 1.5× the output rate; batch and flex run 50% below standard; fast mode runs at 2× standard.

Qwen4-Max Preview has none of that. It is metered in Credits against a subscription, and Alibaba does not publish a credits-per-token rate. Personal Edition tiers run $6 for 20,000 credits, $10 for 45,000, $18 for 80,000 and $43 for 200,000, with top-up bundles at $15 per 20,000. During the preview, consumption runs "as low as 20% of the standard rate", and calls between 22:00 and 08:00 UTC+8 get a further discount landing at 8% of standard. Credits are consumed at a rate the vendor describes as dynamically set by model type, token usage, thinking mode and tool calls.

So the honest statement is not "Qwen is cheaper" or "OpenAI is cheaper". It is that one model can be costed before you run it and the other can only be costed after, from a console you may not have access to yet. For a volume workload that distinction outranks any per-token number on this page.

Where the preview actually beats the rate card

Three capabilities are worth more than a price row, at least for a fortnight.

Video input is the first. GPT-6.1 Sol's documented input modalities are text, image and file. Qwen4-Max Preview accepts text, images and video, and Alibaba's page names screenshot analysis, document understanding and video analysis as the intended uses. If your evaluation is about understanding a screen recording, a product demo, or a frame sequence, the comparison has a winner before any score exists.

The second is the output ceiling already mentioned — a legitimate advantage for long-form generation, and the one place where the numbers are far enough apart to matter without a benchmark.

The third is the price of being wrong. A Token Plan subscription caps spend at the subscription. A metered model caps nothing, and a badly configured agent loop on a $10-per-million output rate can produce a memorable invoice in an afternoon. For an experiment you are not sure about, a ceiling is worth paying a little for.

Where GPT-6.1 Sol wins without trying

Everything that requires the model to still exist next month.

There is a contract. There is a published rate, a context and output limit on a public page, an effort ladder from low to xhigh and max, documented migration guidance, and a vendor whose support path does not end on a calendar date. There is also a checkpoint: max_tokens behaves the way you expect, and the failure modes are documented rather than discovered.

The Qwen preview inverts most of that. Access is Token Plan only, pay-as-you-go is explicitly unsupported, the base URL is a dedicated one at token-plan.ap-southeast-1.maas.aliyuncs.com, and the plan's own terms restrict it to "interactive use in coding and agent tools, not automated batch scripts". Alibaba also states that no changes will be made during the preview. Frozen, temporary, and not for batch — that is a precise description of an evaluation vehicle, and a poor description of a dependency.

There is one further trap for anyone porting a request across, and it belongs in this section because it is the kind of thing a stable vendor does not do to you mid-flight. Starting with Qwen4-Max Preview, max_tokens is equivalent to max_completion_tokens and both cover thinking plus the answer, where every earlier Qwen model applied the cap to the answer alone. A request carried over with max_tokens at 4096 now buys 4,096 tokens of thinking and whatever is left. The failure is a truncated response, not an error you can catch.

On measurement, both sides are thinner than the marketing suggests

Artificial Analysis, the third-party board this blog treats as the neutral reference, has no Qwen4 entry at all when read on 11 November 2026 — no index score, no cost-per-task, no speed rows. For a preview that opened that morning, that is expected rather than damning.

The OpenAI side is now measured, and by the same board. Artificial Analysis carries a GPT-6.1 Sol (Max) page: an Intelligence Index of 55 on index revision v5.0.1, rank 12 of 74, $1.27 per index task, 63.7 million output tokens generated while running the index against a board median of 72.8 million, 20.2 minutes per task against a 12.6-minute median, 56 tokens per second, and a time to first answer token of 153.34 seconds. Those are independent figures, not OpenAI's, and they sit next to the vendor line: OpenAI sells this model as near-Astra at a fifth of Astra's token prices, and on the price card the arithmetic checks out exactly — $2 against Astra's $10 input, $10 against Astra's $50 output.

What that leaves is one number, not a comparison. GPT-6.1 Sol has a row; Qwen4-Max Preview does not and may not get one before it retires, so anyone publishing a Qwen4-Max Preview versus GPT-6.1 Sol benchmark this week is publishing their own harness on both sides. The useful thing to ask them is what prompts they used and whether the Qwen side still exists when they try to reproduce it.

Running both without two contracts

Here is the practical asymmetry: GPT-6.1 Sol is a catalogue model and Qwen4-Max Preview is not — for us or for anyone outside Alibaba's Token Plan. On OrcaRouter, openai/gpt-6.1-sol sits alongside 216 other model identifiers behind one key, at the provider's list price with nothing added, which means a change to OpenAI's rate card lands on your bill the day it happens rather than at the next negotiation. The Qwen preview is reachable through the vendor's own API and its chat product, and no routing layer changes that.

The useful pattern for the next fourteen days is to put the preview where it cannot hurt you. A routing configuration that tries the preview first and falls through to a model you have already measured gives you the experiment and the safety net in one call: if the preview truncates because the thinking budget ate the answer, errors on a video input, or simply returns something you would not ship, the fallback answers and the request still succeeds. That is what failover is for, and an unproven endpoint with an expiry date is the clearest case for it you will get this quarter.

The decision, in four lines

If you need a model in production this month, GPT-6.1 Sol is the answer and the preview is not a candidate — not because it is worse, but because on 26 November it will not exist.

If you need to know whether a million-token window plus video input solves a specific problem, the preview is the cheapest way to find out, and the chat tab makes it cheaper still.

If you are running agents with heavy prefix reuse, the cached-input line is the one to model, and only one of these two models lets you model it.

If you are waiting for a score before committing to either, you are waiting on different things: an independent row for Qwen4-Max Preview that may never appear because the model retires first, and a second opinion on GPT-6.1 Sol's 55, which exists but is a single number from a single harness on a board that was revised to v5.0.1 after the September release.

The comparison people will keep making — benchmark against benchmark — cannot be made today. The comparison that can be made is about commitment: a fortnight-long loan of a bigger context and a video decoder, against a metered contract with a cache rate you can plan around. Pick the one that matches how long you need the answer to stay true.