
Qwen4-Max vs GPT-6.1 Sol: The Cheaper Token Loses on the Invoice, and the Boundary Is at 272,000
- GoogleNEWGoogle: Gemini 4 Carbon2026-11-10$2.00 / $10.00 per 1M tokens · 90 tok/s
- AnthropicNEWAnthropic: Claude Fable 5.22026-11-09$8.00 / $40.00 per 1M tokens · 97 tok/s
- orcaNEWOrca: OrcaC2 1.02026-11-0711 tok/s
- OrcaNEWOrcaCyber: GLM 5.3 Abliterated / Uncensored2026-11-04$3.00 / $5.00 per 1M tokens · 88 tok/s
- OrcaNEWOrcaCyber: Hy4 Abliterated / Uncensored2026-11-04$0.80 / $1.60 per 1M tokens · 133 tok/s
- OrcaNEWOrcaCyber: Kimi K3.1 Pro Abliterated / Uncensored2026-11-04$4.00 / $20.00 per 1M tokens · 86 tok/s
- OrcaOrcaCyber: MiniMax M3.1 Flash Abliterated / Uncensored2026-10-29$0.80 / $1.60 per 1M tokens · 125 tok/s
- OrcaOrcaCyber: GLM 5.3 Flash Abliterated / Uncensored2026-10-27$0.80 / $1.60 per 1M tokens · 67 tok/s
- OrcaOrcaCyber: Qwen3.8 Flash Next Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 92 tok/s
- OrcaOrcaCyber: MiMo V2.6 Flash Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 44 tok/s
- MoonshotMoonshotAI: Kimi K3.1 Pro2026-10-2447Intelligence
- OrcaOrcaCyber: DeepSeek V4.1 Flash Abliterated / Uncensored2026-10-23$0.80 / $1.60 per 1M tokens · 193 tok/s
- minimaxMiniMax: M3.1 Flash2026-10-1929Intelligence
- OrcaOrca: OrcaCyber Ultra 1.02026-10-17$5.00 / $10.00 per 1M tokens · 134 tok/s
- deepseekDeepSeek: DeepSeek V4.1 Flash 10172026-10-1733Intelligence
- DeepSeekDeepSeek: DeepSeek V4.1 Flash2026-10-1740Intelligence
- OrcaOrca: OrcaCyber Zero 1.52026-10-10$3.00 / $7.50 per 1M tokens · 126 tok/s
- openaiOpenAI: GPT-6.1 Sol2026-09-2955Intelligence
- anthropicAnthropic: Claude Sonnet 5.52026-09-2855Intelligence
- typesafeTypeSafe: Jev 1.132026-09-24$0.04 / $0.00 per 1M tokens · 185 tok/s
Take the two rate cards at face value and Qwen4-Max wins by a distance. It lists at $1.50 per million input tokens and $4.50 per million output; GPT-6.1 Sol lists at $2.00 and $10.00. On output that is a 2.2× spread in Qwen4-Max's favour, and it is the number every comparison of these two models leads with. Then run both through the same nine evaluations and look at what the runs cost. Artificial Analysis measures $3.27 per Intelligence Index task for Qwen4-Max, generally available since 16 November 2026, against $1.27 for GPT-6.1 Sol, which OpenAI has sold since 29 September 2026. The cheaper token is 2.6× more expensive per finished task. Two mechanisms produce that inversion, and only one of them is discussed — the other sits at 272,000 tokens, where Sol's card changes price mid-request and Qwen4-Max's does not.
Two cards with different shapes, not just different numbers

• Input price — Qwen4-Max $1.50 per million flat vs GPT-6.1 Sol $2.00 up to 272,000 tokens, $4.00 above
• Output price — $4.50 flat vs $10.00 up to 272,000 tokens, $15.00 above
• Cache read — $0.075 per million with writes at $1.875 (5-minute) or $3.00 (1-hour) vs $0.10 per million with writes at $2.50 and $5.00
• Context window — 1,000,000 tokens with a 983,616-token thinking-mode input ceiling vs 1,050,000 tokens
• Max output — 131,072 non-thinking, or 262,144 shared with reasoning vs 128,000
• Reasoning control — three levels (low, high, xhigh), default high vs five (low through max), default medium, always on
• Sampling parameters — temperature, top_p and top_k accepted vs rejected; reasoning must be steered by effort alone
• Input modalities — text, image, video vs text, image, file
The first two rows are where the money is. Qwen4-Max sells one tier from zero to the context limit, deliberately: no long-context surcharge, no batch discount either, because batch inference is not offered on the model at all. GPT-6.1 Sol sells a cheap tier and an expensive one, and the boundary between them is the single most consequential number in this comparison for anyone whose prompts are large.
What crossing 272,000 tokens does to the arithmetic
Below the boundary, a million input tokens on GPT-6.1 Sol costs $2.00 and a million output tokens $10.00. At or above it, the catalogue bills the request at $4.00 in and $15.00 out — a doubling on input and a 50% increase on output. Qwen4-Max charges $1.50 and $4.50 at 200,000 tokens and at 900,000 tokens alike.
Run a document-heavy job through both and the ordering flips depending on where the request lands. A 150,000-token prompt with a 4,000-token answer costs roughly $0.26 on Sol and $0.24 on Qwen4-Max — near enough to a tie that nothing else about the models should be decided on price. The same job at 400,000 input tokens costs $1.65 on Sol at the upper tier against $0.62 on Qwen4-Max: a 2.7× gap, in the opposite direction from the headline output spread and larger than it. Long-context work is the one workload where the Qwen4-Max card is straightforwardly the better buy, and it is the workload this pair is most likely to be pointed at, since both models are sold on 1M-token windows.
Which makes the tier boundary a routing decision rather than a pricing footnote. If a pipeline's requests cluster just under 272,000 tokens, splitting or caching to stay below the line is worth real money; if they sit well above it, the split is pointless and Qwen4-Max is cheaper on input by a factor of nearly three.

The inversion that has nothing to do with price
Verbosity is the other mechanism, and it runs against Qwen4-Max harder than the rate card runs for it. Over the Intelligence Index, Artificial Analysis recorded 63.7 million output tokens from GPT-6.1 Sol — concise, by its measure, against a median of 78.6 million — and 127 million from Qwen4-Max, which it calls very verbose. Twice the output tokens at less than half the price per token nets out to a bill that is 2.6× larger.
This is not a quirk of the evaluation. Alibaba's own launch material makes the same point from the other side: Qwen4-Max's headline claim is that it cuts token use 24% to 35% against its predecessor, Qwen3.8-Max, and 48% to 49% on Terminal-Bench 4.0. Those are real improvements and they are relative. Against GPT-6.1 Sol, the model is still spending roughly two output tokens for every one Sol spends, and the launch post does not claim otherwise.
The practical consequence is that any cost model built by multiplying your token count by the list price will get the sign of this comparison wrong. Model the tokens first. If you already have traffic on a concise model and are considering a switch, expect your output volume to roughly double even as your per-token rate falls by more than half — which on this pair is a net loss.
Speed, where only one side has published anything
GPT-6.1 Sol is slow and it is measured: 23.1 minutes per Intelligence Index task, 168.56 seconds to first answer token, 49 tokens per second output speed on Artificial Analysis's run against OpenAI's endpoint. Our own playground traffic, sampled over the seven days to 16 November, shows a faster median — around 98 output tokens per second with a 1.7-second p50 to first byte — which is what you would expect from a gateway route sitting in front of a different upstream than the one Artificial Analysis used. Both numbers can be true; neither describes the other's path.
Qwen4-Max has no independent speed number at all. Artificial Analysis lists the model's index score, its price and its token totals, and leaves time per task, output rate and time to first token blank. Alibaba publishes token-efficiency claims, which are not latency. So for one side of this comparison you can capacity-plan the full path, and for the other you can capacity-plan the bill and nothing else. Anyone deploying an interactive surface on Qwen4-Max this month is measuring latency for the first time themselves.
Two migration details that will cost you a debugging afternoon
GPT-6.1 Sol's reasoning is always on with five effort levels and a default of medium, and the model rejects temperature and top_p outright — sampling is steered through effort and nothing else. Qwen4-Max accepts the full sampling set, with defaults of 1.0 temperature and 0.95 top_p in thinking mode, 0.7 and 0.8 outside it, plus top_k 20 and a presence penalty of 1.5 when thinking is off. A client migrated from Sol to Qwen4-Max will silently acquire sampling behaviour it never had; the reverse migration will throw.
Effort strings are the mirror image. Sol offers five rungs; Qwen4-Max offers three and folds the rest, mapping minimal to low, medium to high and max to xhigh. A pipeline requesting medium from Qwen4-Max gets high-effort reasoning with a default thinking budget of 131,072 tokens and no error to announce it. A pipeline requesting max gets xhigh and looks correct. Pin the level in the vendor's own vocabulary per route; do not pass an effort string through a model boundary and hope.

Where the two belong, and how to run them side by side
The split that falls out of the numbers: GPT-6.1 Sol for short-prompt, high-volume reasoning where its concision compounds, and Qwen4-Max for requests past 272,000 input tokens, for video input, or for workloads that need sampling control rather than an effort dial.
Running that split is normally the expensive part, and it is the case OrcaRouter is built for on one side. GPT-6.1 Sol is routed at $2.00 and $10.00 per million below the tier boundary and $4.00 and $15.00 above it — the vendor's own rates, passed through with 0% markup, so the tier arithmetic above is what you actually pay rather than an approximation of it — through the same key and the same OpenAI-compatible endpoint as Claude Opus 5.5, Grok 4.5, Qwen3.8-Max and Tencent HY4 Preview. Qwen4-Max is not in the catalogue yet, so that side stays on a direct Alibaba key for now.
Which is still enough to make the comparison real: define the boundary once in a routing rule, send anything under 272,000 tokens to the Sol route and anything above it to Qwen4-Max, and let failover cover the new model while you establish its latency. The routing DSL is what makes that a rule rather than a service — a condition on request size, two destinations, one endpoint, and an automatic fallback if the experimental side misbehaves.
The short version, with its expiry date
Qwen4-Max is the cheaper model per token and the more expensive one per task, and the reason is that it writes twice as much. GPT-6.1 Sol is the cheaper model per task and the more expensive one per long request, and the reason is a tier boundary at 272,000 tokens that its competitor does not have. Neither of those facts is about intelligence, where the two sit twelve index points apart — 43 against 55 at Sol's max effort — in Sol's favour.
The comparison has an expiry date. Alibaba's launch post says smaller Qwen4 models are coming to open source, and a per-token card on a 2.4-trillion-parameter model is exactly the kind of thing that gets revised once the first month of agent traffic lands. Until then, decide on token volume and request size, not on the output rate.
Compared in this article1
Detected from this article · Benchmarks: Artificial Analysis · updated daily
