APIMODELSAPIMODELS

Qwen 4 Max vs Gemini 4 Carbon

TL;DR Gemini 4 Carbon scores higher on independent benchmarks, and Qwen 4 Max was the more dependable of the two in our own tests. Artificial Analysis scores Carbon at 60 and Qwen 4 Max (Alibaba’s Qwen4-Max, released 16 November 2026) at 43, and Vals has Carbon ahead 70.45% to 63.20%. In our seven same-prompt tasks Qwen 4 Max got all 93 checks right and Carbon 89: Carbon refunded an item the policy ruled out in three runs out of three, and miscounted two figures in a long log. Carbon accepts audio and PDF as well as video; Qwen does not. List prices are $1.50 / $4.50 per million tokens for Qwen 4 Max and $2 / $10 for Carbon until 30 November 2026, then $4 / $20. Both run on one apimodels.app key.

Qwen 4 Max and Gemini 4 Carbon reached developers six days apart: Google opened Carbon to all paid Gemini API users on 10 November 2026, and Alibaba opened Qwen 4 Max for pay-as-you-go use on 16 November under the model ID qwen4-max. Both are the deep-thinking model in their family, both have a one-million-token window, and both take video.

There is less outside evidence for this pair than for most. Alibaba’s launch chart does not include Carbon, and at the time of writing nobody has published a same-task test of the two. So this page leans on two things: the independent labs that have scored both, and seven tasks we ran ourselves on both models with identical prompts.

We run apimodels.app, which sells both models through one API key, so we have no reason to favour either. The benchmark scores come from Artificial Analysis, Vals AI and OpenRouter, with links at the end. The seven same-prompt tests are ours, run on 17 November 2026 through our own API with each model’s default thinking setting.

Qwen 4 Max

Steadier on rules, stable price

  • ·93 of 93 checks in our seven same-prompt tests (Carbon: 89)
  • ·Followed the refund policy in 3 runs of 3; Carbon broke it in 3 of 3
  • ·$1.50 / $4.50 per 1M tokens, no scheduled change
  • ·Text, image and video input; up to 131K output tokens
Qwen 4 Max price and API →

Gemini 4 Carbon

Higher benchmark scores, more input types

  • ·Artificial Analysis Intelligence Index 60 (Qwen 4 Max: 43)
  • ·Vals Index 70.45% vs 63.20%; Terminal-Bench 4.0 65.2% vs 48.5%
  • ·$1.10 / $5.50 per 1M tokens on apimodels.app until 30 Nov, then $2.20 / $11.00
  • ·Text, image, video, audio and PDF input
Gemini 4 Carbon price and API →

Pick Gemini 4 Carbon for the hardest reasoning and research problems, terminal work, and audio or PDF input. Pick Qwen 4 Max when the model has to follow written rules exactly, for legal and document agents, and for a price that does not change on 1 December.

Specs and prices side by side

The headline difference is input types. Both read text, images and video, but only Carbon also takes audio and PDF files directly. If your pipeline starts from recordings or scanned documents, that decides it before any benchmark does.

The second difference is the calendar. Google’s launch price for Carbon, $2 / $10 per million tokens, runs until 30 November 2026 and doubles to $4 / $20 on 1 December. Qwen 4 Max is $1.50 / $4.50 with no change announced. Today Carbon’s input is the lower of the two on apimodels.app; from December Qwen is lower on both sides.

Qwen 4 MaxGemini 4 Carbon
Maker, open to developersAlibaba, 16 Nov 2026Google, 10 Nov 2026
List price per 1M tokens (input / output)$1.50 / $4.50$2 / $10 until 30 Nov; $4 / $20 from 1 Dec
Cached input, list$0.075$0.10 until 30 Nov; $0.20 from 1 Dec
Price on apimodels.app$1.50 / $4.50$1.10 / $5.50 until 30 Nov; $2.20 / $11.00 from 1 Dec
Context window1M1M (1,048,576)
Max output131K (262K with thinking)65,536 by default
Input typesText, image, videoText, image, video, audio, PDF
Thinking controlOn by default, can be switched offThree levels: low, medium, high

Independent benchmarks

On the tests where an independent lab has run both, Carbon is clearly ahead. Artificial Analysis scores it at 60 on the Intelligence Index, fourth of the models it tracks, and Qwen 4 Max at 43, thirty-eighth of 81. On the Vals Index it is 70.45% against 63.20%, and on Terminal-Bench 4.0 65.2% against 48.5%.

Two caveats on reading that gap. Artificial Analysis ran Carbon at its high thinking level and has tested only one setting for Qwen 4 Max so far. And a higher score costs more to produce: its figure per index task is $5.48 for Carbon and $3.27 for Qwen.

There is one place where a published number favours Qwen: the Briefcase office-work ranking, where Alibaba reports 1686 for Qwen 4 Max and Artificial Analysis measured 1611 for Carbon. The two figures come from different runs, so treat that one as an indication, not a result.

MeasureRun byQwen 4 MaxGemini 4 Carbon
Intelligence Index v5.0.1Artificial Analysis43 (38th of 81)60 (4th)
Cost per index taskArtificial Analysis$3.27$5.48
Vals Index v2.2Vals AI63.20% (11th of 38)70.45%
Terminal-Bench 4.0Vals AI / Artificial Analysis48.5%65.2%
Output speed, live medianOpenRouter30 tokens/s55 tokens/s
Latency to first token, live medianOpenRouter3.6 s18.6 s

Our seven same-prompt tests: Qwen 93 of 93, Carbon 89

We wrote seven tasks modelled on what developers most often send through our API: writing a function, fixing a buggy class, a support agent that calls tools and must follow a refund policy, questions over a 1,500-line production log, turning a Chinese group chat into JSON, three counting puzzles, and Chinese product copy with hard constraints. Each has an automatic check, 93 checks in total, verified against reference answers before we ran anything.

On four of the seven the two were identical: both passed all 50 hidden cases on the function, found all five planted bugs, extracted the six orders from the chat with the right dates, and solved all three puzzles. Qwen 4 Max then went on to get the other three right as well. Carbon did not.

The refund task is the one that matters most. A customer asks for money back on a defective keyboard and on a mouse they simply do not like; the order was delivered 46 days earlier, and the policy the agent can look up allows change-of-mind returns only within 30 days. Carbon read the policy, then refunded the mouse anyway and told the customer it was “within 30 days of the October 2nd delivery” on 17 November. We ran it three times and it did the same thing three times. Qwen 4 Max refused the mouse and explained why in all three of its runs.

On the log, Carbon answered in 3.5 seconds and got two of five wrong: it counted seven payment timeouts where there are eight, and named the wrong service as the one with the most errors (the real leader has roughly twice as many as the service it picked). Qwen took 63 seconds and got all five. On the product copy Carbon wrote 221 Chinese characters against a limit of 220.

Carbon’s response time on the night of our run was erratic. Some tasks came back in under 7 seconds; the puzzles took nearly 15 minutes and the function over 7. The whole set took 1,501 seconds on Carbon and 415 on Qwen 4 Max. We would not read that as Carbon’s normal speed, see the speed section, but it is what happened. The bill was close: $0.17 for Carbon at this month’s price and $0.21 for Qwen.

Task (checks)Qwen 4 MaxGemini 4 Carbon
Write a timeline-merge function (50 hidden cases)50/50 · 51 s · $0.01250/50 · 459 s · $0.007
Fix a TTL cache class with 5 planted bugs (10)10/10 · 27 s · $0.00910/10 · 7 s · $0.013
Refund agent with tools and a policy (7)7/7 · 16 s · $0.0066/7 · refunded an excluded item, 3 runs of 3
Five questions over a 1,500-line log (5)5/5 · 63 s · $0.1203/5 · 4 s · $0.086
Chinese group chat to JSON (13)13/13 · 19 s · $0.00413/13 · 62 s · $0.013
Three counting puzzles (3)3/3 · 196 s · $0.0493/3 · 891 s · $0.032
Chinese product copy, five hard constraints (5)5/5 · 43 s · $0.0104/5 · 16 s · $0.016 (one character over)
All seven93/93 · 415 s · $0.2189/93 · 1,501 s · $0.17

What a task costs

Until 30 November the two are close on price. At list, Qwen 4 Max is $1.50 / $4.50 and Carbon $2 / $10; on apimodels.app Carbon is $1.10 / $5.50, so its input is lower than Qwen’s and its output a little higher. In our seven tasks that came to $0.17 on Carbon and $0.21 on Qwen.

From 1 December that changes. Carbon’s list price doubles to $4 / $20 and ours to $2.20 / $11.00, while Qwen 4 Max stays at $1.50 / $4.50. The same seven tasks would then cost about $0.34 on Carbon, against the same $0.21 on Qwen.

Both models think before they answer and bill the thinking as output. On hard benchmark tasks that dominates the cost: Artificial Analysis’s figure per task is $5.48 for Carbon and $3.27 for Qwen 4 Max.

Speed

Carbon writes faster once it starts, and takes longer to start. OpenRouter’s live medians show 55 output tokens per second for Carbon against 30 for Qwen 4 Max, and 18.6 seconds to the first token against 3.6. Artificial Analysis measured 62 tokens per second for Carbon with 74 seconds to the first answer token at high thinking; it has not published a speed for Qwen yet.

On our own platform over the last 30 days, Carbon’s first token arrived after about 27 seconds at the median (production calls, sample size not disclosed). Qwen 4 Max went on sale here on 16 November, so there is no comparable production figure yet; in our tests its first visible word took 15 to 62 seconds outside the puzzles.

Neither is a good fit for a chat box where someone is watching a spinner. If you need either to answer quickly, turn the thinking down: Carbon has a low level, and Qwen 4 Max lets you switch thinking off per request.

Which to use

Hard research, science and multi-step debugging: Gemini 4 Carbon. It is seventeen points ahead on the Artificial Analysis index and seven on the Vals Index.

Agents that act on written rules (refunds, approvals, compliance checks): Qwen 4 Max. It applied the policy correctly every time in our test; Carbon did not in any of its three runs. Whichever you use, check dates and thresholds in code and do not leave them to the model.

Exact counts and lookups over long logs or documents: Qwen 4 Max in our test, 5 of 5 against 3 of 5. If you use Carbon for this, ask for a higher thinking level; it answered ours in under four seconds.

Audio or PDF input: Gemini 4 Carbon. Qwen 4 Max takes text, images and video only.

Long single answers: Qwen 4 Max, with up to 131K output tokens against Carbon’s 65,536 default.

Work you will still be running in December: price it at $4 / $20 for Carbon before you decide.

On apimodels.app switching between them is a change of the model string on the same key, so run a sample of your own tasks on both before committing.

Caveats

Our seven tasks are everyday-difficulty work at default settings, and apart from the refund task, which we repeated three times, each was run once. Four checks out of 93 is a small difference; the refund result is the one we would rely on, because it repeated. The tasks do not measure the hard end, which is where the independent benchmarks put Carbon well ahead.

Qwen 4 Max has been generally available for one day at the time of writing, Artificial Analysis has tested only one thinking setting for it, and no third party has yet compared these two models on the same tasks. Carbon’s slow responses on the night we tested may reflect load at that hour and should not be taken as typical.

Sources

Specs and list prices: Alibaba’s QwenCloud model page and Google AI Studio’s pricing and Gemini 4 Carbon pages, read on 17 November 2026. Independent scores and task costs: Artificial Analysis and Vals AI. Live speed and latency: OpenRouter’s comparison page, read the same day. The seven same-prompt tests are our own, run on 17 November 2026; the refund task was repeated three times on each model.

Quick start

cURL

curl https://api.apimodels.app/v1/chat/completions \
  -H "Authorization: Bearer $APIMODELS_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "qwen4-max",
    "messages": [{"role": "user", "content": "Apply this refund policy to the order below: ..."}]
  }'
# same key, same endpoint: change "model" to "gemini-4-carbon" to run Carbon

FAQ

Is Qwen 4 Max better than Gemini 4 Carbon?

On benchmarks, no: Gemini 4 Carbon scores 60 on the Artificial Analysis Intelligence Index against 43 for Qwen 4 Max, and 70.45% against 63.20% on the Vals Index. In our seven same-prompt tests the result went the other way, 93 of 93 for Qwen 4 Max and 89 of 93 for Carbon, because Carbon broke a refund policy in three runs out of three and miscounted two figures in a long log.

Which costs less, Qwen 4 Max or Gemini 4 Carbon?

Until 30 November 2026 they are close: Qwen 4 Max lists at $1.50 / $4.50 per million tokens and Gemini 4 Carbon at $2 / $10. From 1 December Carbon’s list price doubles to $4 / $20 and Qwen 4 Max becomes the lower-priced model on both input and output. Per Artificial Analysis task the figures are $3.27 for Qwen and $5.48 for Carbon.

Which is faster, Qwen 4 Max or Gemini 4 Carbon?

Gemini 4 Carbon outputs faster, about 55 tokens per second against 30 on OpenRouter’s live medians, but takes longer to produce its first token, 18.6 seconds against 3.6. Both think before answering, so neither suits a chat interface without turning the thinking down.

Can Qwen 4 Max read audio or PDF files?

Not directly. Alibaba lists text, image and video input for Qwen 4 Max. Gemini 4 Carbon accepts text, image, video, audio and PDF.

Which is better for customer-support agents that follow a policy?

In our test, Qwen 4 Max. Given a refund policy with a 30-day window and an order delivered 46 days earlier, Qwen 4 Max refused the out-of-window refund in three runs of three, while Gemini 4 Carbon issued it in three of three. With either model, enforce dates and amount limits in your own code.

Can I use Qwen 4 Max and Gemini 4 Carbon with one API key?

Yes. Both run on one apimodels.app key. Call POST https://api.apimodels.app/v1/chat/completions with model qwen4-max or gemini-4-carbon.

Related

Мы на
APIMODELS on Product HuntAPIMODELS featured on ToolifyAPIMODELS featured on There's An AI For That