A generated hero card for a comparison piece, headed 'Qwen4-Max vs GPT-6.1 Sol Pro' with the subtitle 'Pro is a billing mode, and it has a number to beat'. A highlighted figure reads 'Break-even for pro mode: 2.6X', with chips reading 'Qwen4-Max: $3.27 per task' and 'GPT-6.1 Sol at max: $1.27 per task'. A footer reads 'Artificial Analysis and OpenAI documentation read 17 November 2026'. The OrcaRouter mark is composited bottom right.
Engineering & Research

Qwen4-Max vs GPT-6.1 Sol Pro: Pro Is a Billing Mode, and It Has a Number to Beat

Author

jinhao.song

Date Published

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

GPT-6.1 Sol Pro does not have a price, because it does not have an identifier. It is GPT-6.1 Sol — released on 29 September 2026 at $2 per million input and $10 per million output tokens — called with reasoning.mode: "pro" in the Responses API, a setting O​penAI documents as making the model do more work and billing that work at the same standard token rates. Qwen4-Max, generally available since 16 November 2026 at $1.50 and $4.50, is the cheaper model per token by a factor of 2.2 on output. It is also, on the only independent measurement either model has, the more expensive one per finished task: $3.27 against $1.27 for GPT-6.1 Sol at its max effort setting. That inversion — Qwen4-Max cheaper per token and dearer per answer — is the whole comparison, and pro mode is the one variable that could move it in either direction without anybody being able to measure which.

Pro mode is a multiplier with no published coefficient

Two paragraphs of vendor documentation are all that exist, and they are worth reading closely because every "Sol Pro" comparison on the internet is built on them plus inference.

• What it is — a reasoning.mode value of pro, set alongside reasoning.effort, in the Responses API. The other value is standard, which is the default.

• What it does — per O​penAI, it "aggregates the model work performed to produce the final answer". It is independent of effort: the mode picks how much work happens, effort picks how hard the model thinks within it.

• What it costs — the aggregated tokens, billed at the model's standard rates. No surcharge, no separate meter.

• What it does not have — a coefficient. O​penAI publishes no token multiplier for pro mode, and the neutral leaderboard carries no pro-mode row: every GPT-6.1 Sol measurement on Artificial Analysis is a reasoning.effort setting from low to max, all in standard mode.

Which means one thing follows with certainty and one does not. Certainly: pro mode can only raise GPT-6.1 Sol's cost per task, never lower it, because it adds billed tokens at the same rate. Not certainly: by how much. A pro-mode run that aggregates 1.8× the work is a different product from one that aggregates 4×, and nobody outside O​penAI has published the ratio.

The threshold pro mode has to clear

That unknown is not a reason to shrug. It converts into a single number, because Qwen4-Max's cost per task is already measured and so is Sol's.

• At Sol's max effort — $1.27 per Intelligence Index task against Qwen4-Max's $3.27. Pro mode would have to multiply Sol's token spend by 2.57× before the two models cost the same per task.

• At Sol's xhigh effort — $0.76 against $3.27. The threshold is 4.3×.

• At Sol's high effort — $0.59 against $3.27. The threshold is 5.5×.

• On intelligence, for scale — Sol scores 54.6 on the index at max, 52.5 at xhigh, 50.3 at high. Qwen4-Max scores 42.9. Pro mode does not change a model's intelligence score in the sense the index measures it, but it changes the answer quality on hard tasks, which is the only reason to pay for it.

So the question a buyer actually faces is not "which is cheaper". It is: does a pro-mode Sol run beat Qwen4-Max by more than ten index points' worth of quality, while staying under 2.6× to 5.5× its standard-mode token spend? Every published signal says the multiplier is real but modest — the vendor frames pro mode as the setting for "difficult tasks that need more model work", not as a different model — and a multiplier under 2.6× would leave Sol cheaper per task than Qwen4-Max at every effort level. That is an inference, clearly labelled as one, and it is the most defensible thing anyone can say until O​penAI or an independent evaluator publishes the ratio.

A generated two-column scoreboard titled 'Qwen4-Max vs GPT-6.1 Sol — the scoreboard'. The Qwen4-Max column reads Intelligence Index 43, Cost per task $3.27, Output tokens 127M, Price $1.50 / $4.50, Effort curve one point, Speed not measured. The GPT-6.1 Sol column reads Intelligence Index 55, Cost per task $1.27, Output tokens 63.7M, Price $2.00 / $10.00, Effort curve five points, Speed 49 tok/s. A footer reads 'Both columns per Artificial Analysis, 17 November 2026. Pro mode has no published multiplier.'

Why the cheap model is the expensive one

The per-token card and the per-task bill disagree by a factor of nearly six, and the reason is volume, not price. This is worth spelling out because it is the single most common error in comparisons of these two models.

• Output rate — Qwen4-Max $4.50 per million against Sol's $10.00. Qwen4-Max is 2.2× cheaper per output token.

• Output volume on the index run — 127 million tokens for Qwen4-Max, 63.7 million for Sol at max effort. Qwen4-Max writes 2.0× as much.

• Verbosity rank — Qwen4-Max sits 65th of 80 on the tracker's scale, squarely in the "very verbose" band; Sol at max sits 35th, near the middle.

• Realised cost per task — $3.27 against $1.27. Qwen4-Max is 2.6× dearer, not 2.2× cheaper.

Those three ratios do not reconcile into the fourth, and the missing term is input volume, which the public page does not break out per model. That is exactly why the per-task figure should be treated as the measurement and the ratios as the explanation: a rate card describes what a token costs, and the index describes what a model spends, and only the second one arrives on an invoice at the end of the month. A​libaba's own launch numbers make the same point in the other direction — average cost per task on its Work360 benchmark falling from $4.43 to $1.67 is a claim about that benchmark's token mix, and it is a vendor figure on a vendor benchmark.

One more axis, because it is the one that decides agent feasibility rather than cost. Sol at max effort took 1,383 seconds of measured time per index task — twenty-three minutes — with a 168-second time to first token. Qwen4-Max has no time-per-task figure at all; the tracker's speed fields are empty for it. Our own playground's seven-day record for the O​penAI route shows a 1.7-second median first token and 98 output tokens per second, which is standard-mode, medium-effort traffic, not pro. Nobody has published how long a pro-mode call takes, and the honest planning assumption is the max-effort figure with error bars pointing up.

A headless-browser screenshot of the Artificial Analysis model page for GPT-6.1 Sol, captured 17 November 2026 in English, showing the OpenAI vendor label, a Proprietary model tag, 'Released September 2026', the heading 'GPT-6.1 Sol (Max)', an Intelligence Index of 55 ranked 12 of 81, a time to first token of 168.565 s, 49 output tokens per second, 63.7M generated tokens against a 78.6M median with a verbosity rank of 35 of 80, a 1M-token context window, and the cost block reading In $2.00, Out $10.00 and $1.27 per task.

Where the mode belongs: the routing layer, not the code

Because pro is a request parameter and not a model, the decision to use it is a per-call decision, and per-call decisions made in application code age badly. The version that does not age is to put it in the route: a pro-mode configuration pinned to the deep-review path — the one that runs once per pull request, per contract, per incident — and a standard-mode configuration on everything interactive, both reachable from the same key.

OrcaRouter routes openai/gpt-6.1-sol at OpenAI's own $2.00 and $10.00 with 0% markup, including the 272,000-token tier where the rates double to $4.00 and $15.00, so a pro-mode experiment run through us costs what it costs at the vendor and nothing more. Qwen4-Max is not on the catalogue; a request for a qwen4 identifier returns not-found, and the newest Qwen flagship we serve is qwen/qwen3.8-max at $2.00 and $6.00. The routing DSL is where the two halves meet: a rule that sends long-horizon agent work to one model and interactive turns to the other, with automatic failover underneath so a pro-mode run that stalls does not take the request down with it. That is a configuration change, and it is reversible in a way that a hardcoded model string is not.

A headless-browser screenshot of the OrcaRouter model page for openai/gpt-6.1-sol, captured 17 November 2026 in English, showing the GPT-6.1 Sol heading, the model id openai/gpt-6.1-sol, attribution to OpenAI dated 2026-09-29, vision, tools, JSON and reasoning badges, a 1M-token context window with a 128K maximum output, a p50 time to first token of 170 s, a Zero Data Retention note, and the rate block reading $2.00 input and $10.00 output.

What would change the answer

Three things, in order of how likely they are to arrive first.

The first is a pro-mode row on an independent board. If a pro configuration appears next to the five effort levels with its own cost per task, the 2.6× threshold stops being arithmetic on a guess and becomes a measurement, and this page gets rewritten around it. Until then, any article quoting a "GPT-6.1 Sol Pro score" is quoting the base model's standard-mode row with a mode name attached.

The second is a Qwen4-Max effort curve. The model is two days old and has exactly one measured operating point, which means its 42.9 could be a floor or a ceiling. A​libaba's launch material references High and Xhigh settings on its own benchmarks, so the curve exists; it just has not been walked independently yet.

The third is Qwen4-Max getting a speed row. A model with no throughput figure is a model you cannot staff an agent around, and the absence is currently the strongest argument for the incumbent in any workload with a human waiting.

Until one of those lands, the decision is legible. If your workload is interactive and cost-sensitive, GPT-6.1 Sol at medium — $0.36 per task, index 46.4, five-second first token — beats Qwen4-Max on every measured axis including price, which is not the conclusion the token card advertises. If your workload is a long-horizon agent where a 1M-token window, native vision and a 95% cache discount matter more than the index, Qwen4-Max is the cheaper way to buy context, and its bill will be decided by cache hit rate rather than by effort. And if you are weighing pro mode, you are weighing a 2.6× budget on a capability nobody has measured — which is a fine thing to test on one route and a poor thing to assume on a production path.