Qwen4-Max Charges 55% Less Per Output Token Than GPT-6.2 Sol, Then Costs 3.2x More Per Task: $3.27 to $1.01
Alibaba’s Qwen4-Max, generally available since November 16, lists at $1.50 / $4.50 per million tokens to GPT-6.2 Sol’s $2 / $10. On the Artificial Analysis Intelligence Index, it costs $3.27 a task to Sol’s $1.01, takes 29.3 minutes to 9.8, and scores 43 to 57. Vals AI’s Terminal-Bench 4.0 tells the same story in a different harness.
If you pick a model for an app or an agent by the price sheet, Qwen4-Max looks like the deal of the month. Its output tokens cost 55% less than GPT-6.2 Sol’s and its input tokens 25% less. It has a 1M-token context, a flat price across that whole window, and image and video input.
The bill you pay isn’t per token, though. It’s per finished job: tokens read, times tokens written, times how many turns the model needs. On that measure the cheaper model is the expensive one, by 3.2x.
The price sheet
List prices from each lab’s own page, per million tokens, standard tier.
| Model | Input | Cached input | Output | Context |
|---|---|---|---|---|
| Qwen4-Max (Singapore, International) | $1.50 | — | $4.50 | 1M |
| Qwen4-Max (Global regions, Beijing) | $1.413 | — | $4.24 | 1M |
| GPT-6.2 Sol | $2.00 | $0.10 | $10.00 | — |
| Gemini 4 Hydrogen | $1.50 | $0.075 | $7.50 | — |
| Gemini 4 Neon | $3.00 | $0.15 | $15.00 | — |
Alibaba also lists a Batch File API in Beijing at $0.707 / $2.12, and 1 million free tokens for 90 days on Singapore pay-as-you-go. Thinking is on by default at effort high; the other levels are low and xhigh, and the API maps max to xhigh. Maximum output is 262,144 tokens with thinking and 131,072 without.
Per task: the Intelligence Index
Artificial Analysis runs every model through the same 9 evaluations (Intelligence Index v5.0.1) and publishes what one task cost at list price, how long it took and how many tokens the model used. Settings are the ones Artificial Analysis lists for each score.
| Model | Score | Cost / task | Time / task | Output tokens |
|---|---|---|---|---|
| Gemini 4 Neon (xhigh) | 61 | $6.44 | 21.5 min | 283M |
| GPT-6.2 Sol (max) | 57 | $1.01 | 9.8 min | 60.9M |
| GPT-6.1 Sol (max) | 55 | $1.27 | 25.7 min | 63.7M |
| Gemini 4 Hydrogen (xhigh) | 54 | $3.16 | 15.5 min | 304M |
| Kimi K3.1 Pro | 46 | $3.68 | 39.2 min | 113M |
| Qwen4-Max | 43 | $3.27 | 29.3 min | 127M |
| Qwen3.8 Max (0902, xhigh) | 34 | $10.02 | 48.0 min | 150M |
GPT-6.2 Sol scores 14 points higher than Qwen4-Max, finishes a task about 3x faster and costs $1.01 to $3.27. Artificial Analysis ranks Qwen4-Max #51 of 95 on intelligence and #87 of 94 on performance, with a time to first answer token of 123.83 seconds.
Why the cheaper model costs more
The tokens. On the Intelligence Index run, Qwen4-Max read 9.9 billion input tokens to GPT-6.2 Sol’s 2.22 billion, about 4.5x as many (our arithmetic), and wrote 127 million output tokens to Sol’s 60.9 million, about 2.1x. Artificial Analysis puts Qwen4-Max at #78 of 94 for verbosity.
Input tokens that large usually mean more turns: every extra step of an agent loop re-sends the conversation so far. A 25% discount per input token doesn’t survive 4.5x the input. Caching helps both models (Artificial Analysis lists a 95% cache discount for Qwen4-Max), but it shrinks the bill of a model that takes more turns; it doesn’t remove the turns.
Speed adds a second cost. Qwen4-Max writes 60 tokens a second; Sol takes 9.8 minutes a task to Qwen’s 29.3. For a chat feature, a 2-minute wait before the first answer token is a product problem, whatever it costs.
A second opinion: Terminal-Bench 4.0
Vals AI runs Terminal-Bench 4.0 with the same agent harness (Mini-SWE-agent) for every model and reports the cost of each test. It’s a closer match to what a coding agent does than a general index.
| Rank | Model | Accuracy | Cost / test | Duration / test |
|---|---|---|---|---|
| 3 | Claude Sonnet 5.5 | 64.14% | $10.75 | 1h 22m |
| 4 | Kimi K3.1 Pro | 63.64% | $4.57 | 49m 21s |
| 7 | Gemini 4 Neon | 61.62% | $6.42 | 36m 06s |
| 8 | GPT-6.2 Sol | 60.61% | $1.13 | 23m 21s |
| 13 | GPT-6.1 Sol | 55.05% | $1.72 | 45m 12s |
| 14 | Gemini 4 Hydrogen | 54.04% | $7.56 | 30m 40s |
| 18 | Qwen4-Max | 48.48% | $3.90 | 1h 15m |
| 24 | Qwen3.8 Max (0902) | 38.89% | $12.15 | 2h 23m |
Same order: Qwen4-Max costs about 3.5x what GPT-6.2 Sol does per test (our arithmetic) and scores 12 points lower. Two independent boards, two harnesses, one answer.
What Qwen4-Max did get right
Against its own predecessor, Qwen4-Max is a big step. On the Intelligence Index it scores 43 to Qwen3.8 Max’s 34 and costs $3.27 a task to $10.02, about a third. On Terminal-Bench 4.0 it gains 9.6 points and cuts the cost per test from $12.15 to $3.90 and the time from 2 hours 23 minutes to 1 hour 15 minutes. If you already run Qwen3.8 Max, switching is an easy win.
Vals AI also has it at 63.20% on the Vals Index (cost per test $9.33), 7th of 44 on HLAB-Vals-1.1 at 21.25%, and 9th of 46 on Finance Agent v2.1 at 66.68%. It isn’t a weak model; it’s a mid-pack model with a cheap price sheet and an expensive habit.
What this means if you build with AI
Compare cost per task, not cost per token. Run 20 of your own jobs through each candidate and divide the bill by the jobs that passed. The price sheet is an input to that number, not the number.
Cheap agent loops: GPT-6.2 Sol, or GPT-6.1 Sol. On both boards Sol is the cheapest model per finished task in these tables. On MegaBrain, GPT-6.2 Sol isn’t in the catalog yet; GPT-6.1 Sol is, at 55 on the index for $1.27 a task.
Coding agents where the score matters most: Claude Sonnet 5.5 and Kimi K3.1 Pro sit within a point of each other on Terminal-Bench 4.0, and Kimi costs $4.57 a test to Sonnet’s $10.75. Put Sonnet 5.5 in Claude Code (the Claude Code setup guide shows the config on a MegaBrain key) and try Kimi K3.1 Pro in OpenCode for the cheaper runs (the OpenCode setup guide).
Qwen4-Max: batch work where latency doesn’t matter. Long documents, image and video input, overnight jobs: the 1M flat-priced context and the Batch price are where it earns its place. Measure the tokens per job before you commit.
Keep the model name in config. New models land every few weeks, so switching should be a one-line change. Our Gemini 4 Hydrogen cost-per-task post found the same gap between price sheet and bill from the other side.
Sources
Pages opened November 21, 2026. Every number above comes from these pages, or from our arithmetic on them where we say so.
- Alibaba Cloud Model Studio — Qwen4-Max (prices per region, Batch prices, context, output limits, reasoning effort levels and default)
- OpenAI — Introducing GPT-6.2 Astra and Sol (API prices)
- Google AI Studio — Gemini API pricing (Gemini 4 Neon and Gemini 4 Hydrogen)
- Artificial Analysis — Qwen4 Max (Intelligence Index v5.0.1, cost per task, tokens, speed)
- Artificial Analysis — GPT-6.2 Sol
- Artificial Analysis — GPT-6.1 Sol
- Artificial Analysis — Gemini 4 Neon
- Artificial Analysis — Gemini 4 Hydrogen
- Artificial Analysis — Kimi K3.1 Pro
- Artificial Analysis — Qwen3.8 Max (0902)
- Vals AI — Terminal-Bench 4.0 leaderboard (accuracy, cost per test, duration)
- Vals AI — Qwen 4 Max model page (Vals Index and per-benchmark ranks)
Full disclosure: MegaBrain sells pay-as-you-go API access to image, video and text models, including GPT-6.1 Sol, Kimi K3.1 Pro, Claude Sonnet 5.5 and Qwen3.8 Max. MegaBrain is not affiliated with Alibaba, OpenAI, Google, Moonshot AI, Anthropic, Artificial Analysis or Vals AI.
MegaBrain for Mac · an AI coworker
Give MegaBrain the task, not the steps.
It works on your files, writes documents, spreadsheets and decks, connects to your team’s tools and asks before anything risky. Free to start.
Newsletter
Stay in the loop
Get our latest AI research and guides. No spam, unsubscribe anytime.