A generated title card reading 'Qwen4-Max vs GPT-6 Sol' with the subtitle 'The cheaper token that costs more per answer' and three figure cards reading '43 vs 44 on the index', '$3.27 vs $2.47 per task' and '$4.50 vs $10.00 output'. A footer reads 'Scores and per-task cost per Artificial Analysis Intelligence Index v5.0.1; prices per each vendor's rate card.' Flat vector styling on a white background with blue-and-cyan gradient accents, the OrcaRouter mark composited bottom right.
Guides & Insights

Qwen4-Max vs GPT-6 Sol: The Cheaper Token That Costs More Per Answer

Author

jinhao.song

Date Published

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

Qwen4-Max lists at less than half of GPT-6 Sol on output — $4.50 per million tokens against $10.00 — and it still costs more to finish a task. Artificial Analysis measures $3.27 per Intelligence Index task for Qwen4-Max, generally available since 16 November 2026, against $2.47 for GPT-6 Sol, which OpenAI has sold since 22 September 2026. The scores are one point apart, 43 to 44. That inversion is the entire comparison, and it is not a rounding artefact: it is what happens when a model writes 127 million output tokens to complete an evaluation that the other model finishes in 65.1 million. Qwen4-Max is the cheaper model by the unit and the dearer model by the job, and only one of those two facts survives contact with a budget.

The inversion, with the arithmetic shown

Both models run reasoning by default and both vendors price thinking as output. The difference is how much of it there is.

• Output volume on the index run — Qwen4-Max generated 127 million output tokens, which the tracker ranks 65th of 80 on verbosity and describes as very verbose against a median of 78.6 million. GPT-6 Sol generated 65.1 million, described as fairly concise against the same median.

• Input volume — 9.9 billion tokens for Qwen4-Max against 4.75 billion for GPT-6 Sol. Qwen's output-to-input ratio is 1.284%; Sol's is 1.371%, so the shape of the traffic is similar and the scale is not.

• Cost per completed task — $3.27 against $2.47. The nominally cheaper model is 32% more expensive per task.

• Index score — 43 against 44, ranks 38 and 34 of 81 models respectively, on the same revision read on 16 November 2026.

What this measures is not waste in the ordinary sense. A model that reasons at length and restates its working is doing something the vendor considers a feature, and Alibaba's own launch numbers are built on exactly that behaviour being worth paying for: it claims 24% to 35% fewer tokens than Qwen3.8-Max on its Work360 benchmark, which is a claim about relative verbosity against its predecessor and says nothing about how it compares to a model from a different lab. The absolute number is the one that lands on an invoice, and the absolute number is high.

Where the inversion disappears is at fixed output. Constrain the answer with a schema, a max-token ceiling or a strict format and the verbosity advantage evaporates, because the tokens being saved are the ones the model chose to emit. Under a hard output cap, Qwen4-Max's 2.2× cheaper output rate applies to roughly the same token count and the saving is real. Left to write what it likes, it spends the discount.

A generated two-column scoreboard titled 'Qwen4-Max vs GPT-6 Sol — the scoreboard'. The Qwen4-Max column reads Output price $4.50 per million, AA Index 43 rank 38 of 81, Price per index task $3.27, Output tokens on AA run 127M, Above 272K context price unchanged, Speed on AA not yet measured. The GPT-6 Sol column reads Output price $10.00 per million, AA Index 44 rank 34 of 81, Price per index task $2.47, Output tokens on AA run 65.1M, Above 272K context $4.00 / $15.00, Speed on AA 125 tokens per second. A footer reads 'Index and per-task cost per Artificial Analysis v5.0.1; prices per vendor cards.'

Past 272,000 tokens, the price relationship flips outright

GPT-6 Sol's rate card has a cliff in it and Qwen4-Max's does not. This is the one dimension where the cheap model is not merely competitive but dramatically cheaper, and it is easy to miss because both models advertise a million-token window.

• GPT-6 Sol, up to 272,000 input tokens — $2.00 input, $0.20 cached input, $2.50 cache write, $10.00 output.

• GPT-6 Sol, above 272,000 input tokens — $4.00 input, $0.40 cached input, $5.00 cache write, $15.00 output. OpenAI's own model page states the rule plainly: every token in the request is repriced, not just the excess. A 300,000-token prompt bills all 300,000 at the long-context rate.

• Qwen4-Max — $1.50 input and $4.50 output flat across the full 1,000,000-token window, with no tiering at 32K, 128K or anywhere else. Singapore rates; the other five regions are $1.413 and $4.24.

• The gap at 300,000 tokens — $4.00 against $1.50 on input is 2.7×, and $15.00 against $4.50 on output is 3.3×. Add Qwen's batch path in Beijing, where Batch File bills $0.707 and $2.12, and the ratio widens further — though that path exists in one region only.

• The practical consequence — a document-analysis pipeline that routinely crosses 272K tokens is not choosing between two similar prices. It is choosing between a rate that doubles at a boundary its requests cross by accident and a rate that does not move at all. Splitting oversized requests into sub-threshold chunks recovers Sol's price, but it is a change to the pipeline, and it costs you the single-request context that was the reason to buy a million-token model.

This is the strongest case in the piece for Qwen4-Max, and it is narrower than the launch coverage implies: it is a long-context case, not a general one.

What each vendor actually claims, and what stands behind it

Alibaba's launch material for Qwen4-Max is unusually specific and entirely self-administered. The benchmark it leads with, Work360, is one the vendor built: a taxonomy of 26 Tier-1 domains, 182 Tier-2 subdomains and 721 Tier-3 fields, on which Qwen4-Max scores 52.69 against 47.74 for Qwen3.8-Max. Alongside it the company reports token reductions of roughly 24% and 35% on Work360 at xhigh and high effort, 48% and 49% on Terminal-Bench 4.0, an average cost per task falling from $4.43 to $1.67, a Last Secure Code Bench function-level pass rate rising from 22.0% to 31.8%, and code that is secure at function, file and repository granularity at once rising from 8% to 14.7%. Every one of those is vendor-reported and none has been reproduced outside Alibaba.

OpenAI's claims for GPT-6 Sol are shaped differently: they are parity-at-lower-cost arguments against named competitors rather than absolute scores. The launch material reported a FrontierCode 1.1 result of 49.3% at maximum effort for $2.14 per task against 50.3% for Claude Fable 5.1 at $12.83, and a DeepSWE v1.1 result of 68.8% within 1.1 points of Claude Fable 5's 69.9%. Read those sentences for their shape: each pairs a near-tie on score with a fraction of the cost. That is a claim about efficiency, not about capability, and it is a weaker claim than the one Alibaba is making about Work360.

The independent tracker is where the two can finally be compared, and it is the only source in this piece that measured both models with one harness. It puts them one point apart and reverses the price ranking. It also leaves one column empty: GPT-6 Sol has a measured 125 output tokens per second, 35.11 seconds to first answer token and 7.2 minutes per index task. Qwen4-Max has none of the three — the performance card reads N/A. A model that is 32% dearer per task with no published speed is not a model whose cost advantage can be priced at all yet, because the missing half of the bill is time.

A headless-browser screenshot of the Artificial Analysis model page for Qwen4 Max, captured 16 November 2026 in English, showing the Alibaba proprietary-model label, Released November 2026, an Intelligence Index of 43 ranked of 81, the cost block reading In $1.50, Out $4.50, Cache 95% and $3.27 per task, a verbosity reading, and the summary line noting text and image input with a 984k-token context window.

The architecture claim underneath the token numbers

Qwen4-Max is a 2.4-trillion-parameter sparse mixture-of-experts with 71 billion parameters active per token, plus 328 billion N-gram embedding parameters that are looked up rather than multiplied. Three Gated DeltaNet layers compress history into fixed-size state for each Qwen Sparse Attention layer, which selects relevant context at micro-block granularity through a lightweight indexer; four parallel residual branches with dynamic read and write gates stabilise the stack. Alibaba reports no loss spikes across the 2.4T-parameter run and 5% to 13% lower bits per byte than Qwen3.8-Max-Base while activating 25% fewer parameters — with the honest caveat, printed by the vendor, that two of the compared checkpoints are post-trained, which weakens the base-model framing.

It is a plausible account of why long-context inference is cheaper per token. It is not an account of why a request needs 1.95× the output tokens of a competitor, and nothing in the launch material addresses that. Efficiency per token and efficiency per answer are different quantities, and this model improves the first while losing the second.

Where the two live, and what a mixed workload looks like

GPT-6 Sol is routed on OrcaRouter as openai/gpt-6-sol at OpenAI's list price with 0% markup — $2.00 and $10.00 per million tokens under 272K, $4.00 and $15.00 above it, with the 272K repricing rule applied the way OpenAI applies it. One API for 200-plus models means the long-context question above becomes a routing decision rather than a second contract: a pipeline can send its sub-272K requests to Sol and its oversized ones elsewhere without standing up a second SDK, and a vendor price change lands here the same day it lands on OpenAI's card.

Qwen4-Max is not in the OrcaRouter catalogue as of 16 November 2026. It is reachable through Alibaba's own Model Studio API and several third-party platforms. That makes the honest experiment asymmetric and cheap in one direction only: the Sol arm needs no new integration, the Qwen arm does, and the comparison you actually want — cost per completed task on your own traffic, at your own output lengths — is exactly the measurement that would settle whether the $2.47-versus-$3.27 inversion holds outside a benchmark harness. Run it on a shadow slice, not on the production path, and let your own token counts decide which of the two index runs is the misleading one.

A headless-browser screenshot of the OrcaRouter model page for openai/gpt-6-sol, captured 16 November 2026 in English, showing the GPT-6 Sol heading and identifier, the OpenAI attribution dated 2026-09-22, vision, tools, JSON and reasoning badges, a description placing it below GPT-6 Astra and above GPT-6 Luna, an OpenAI-compatible /v1/chat/completions endpoint, and the price block reading $2.00 input and $10.00 output per million tokens with a 4.54 s p50 time to first token.

What would make this a different article

Two measurements. The first is Qwen4-Max's output volume on a workload with a hard output limit; if the 127 million tokens is a property of the evaluation rather than of the model, the inversion is an artefact and the $4.50 rate is a straight saving. The second is speed. GPT-6 Sol burns 7.2 minutes per index task and 35 seconds to first token, which are unremarkable numbers; if Qwen4-Max is materially slower on the same tasks, then a model that is already 32% dearer per answer is also the one that makes you wait longer for it, and no per-token argument rescues that.

Until both exist, the defensible position is narrow and specific. For requests that cross 272,000 tokens, Qwen4-Max is two to three times cheaper and there is no contest. For everything else, GPT-6 Sol is the better buy on the only number that combines price with behaviour, and the fact that its headline rate is 2.2× higher on output is the least informative thing about it.

Compared in this article1

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