Gemini 4 Neon vs Qwen4-Max: Benchmarks, Pricing and Which to Use
A2Agent Team ยท 2026-11-20T00:00:00Z
Alibaba released Qwen4-Max on November 16, 2026, and Google released Gemini 4 Neon two days later. Both are aimed at professional work and long agent tasks, and both have a 1M-token context window, but they're good at different things and priced very differently. Neon is the stronger model on almost every independent measure. Qwen4-Max costs much less and does best on document-heavy work.
How each company positions its model
Qwen4-Max
Qwen4-Max is a sparse mixture-of-experts model with 2.4 trillion total parameters, about 71 billion active per token, and a separate 328-billion-parameter N-gram embedding. It reads text, images and video, and its output budget of up to 262K tokens includes reasoning, which is on by default.
Alibaba calls it an "AI teammate" for work, creation and scientific discovery, and its launch examples are about finished deliverables:
- Live Delivery, a way of running long agent tasks: the model produces a usable first version early and keeps improving it, so you can check progress and correct course while it runs. A companion plugin ships alongside the model.
- In a financial research demo, it used 90 MCP tools to gather and cross-check data, then delivered an Excel workpaper and an editable slide deck where every figure traces to a source.
- In an IPO due-diligence demo, it sorted 198 unorganized files, cross-checked public records and produced an 86-page report with each conclusion cited to its source file.
- In a 48-hour autonomous run with a robotic wet lab, it explored a 2.7-million-molecule space and produced a new data-center coolant molecule that was validated as a real sample.
It can also drive 3D tools such as Blender and Unreal Engine. These are Alibaba's own showcases. Alibaba says the average cost of a task on its Work360 benchmark is 58% lower than with Qwen3.8-Max.
Gemini 4 Neon
Neon is Google's mid-range Gemini 4 model, adapted from Gemini 4 Carbon and placed between Carbon and the lightweight Hydrogen. It reads text, images, audio, video, charts, spreadsheets and PDFs, and has four thinking levels from Low to Extra High.
Google pitches it for three kinds of work. The first is production software engineering: codebase migrations, multi-file refactoring and fixing tests. In Google's Antigravity, Neon can plan a job and hand the parts to several Hydrogen agents running in parallel. The second is security, through Neon for Cyber, a version for vetted defenders that scores 90.5% on a vulnerability-finding test and 70.7% on a patching test. The third is science, where Google says Neon proposes and tests hypotheses on its own instead of just running instructions.
Pricing
| Per 1M tokens | Qwen4-Max | Gemini 4 Neon |
|---|---|---|
| Input | $1.50 | $3.00 |
| Cached input | $0.075 | $0.15 |
| Output | $4.50 | $15.00 |
Neon costs twice as much for input and about 3.3 times as much for output. For a workload of 1 million input tokens and 200,000 output tokens, Qwen4-Max comes to $2.40 and Neon to $6.00.
Benchmarks
Independent rankings
| Qwen4-Max | Gemini 4 Neon | |
|---|---|---|
| Artificial Analysis Intelligence Index | 43 (9th) | 61 at Extra High (3rd) |
| Vals Index | 63.20% (11th) | 70.9% (2nd, per Google) |
| Output speed | Slow and verbose, per testers | About 110 tokens/s |
Neon is well ahead on both indexes. Only GPT-6.2 Astra and Claude Fable 5.2 score higher on Artificial Analysis. Qwen4-Max is the highest-ranked Chinese model on the Vals Index, but it trails Neon by about seven points there.
Cost per task narrows the price gap. On Artificial Analysis's runs, Qwen4-Max cost about $3.27 per task and Neon about $6.44, so Neon was roughly twice as expensive per task, against a much higher score. Neon also took about 21 minutes per task at Extra High, compared with about 36 for Carbon.
Where each one leads
Neon leads on coding agents. Google reports 58.6 on Artificial Analysis's Coding Agent Index, first place, and Artificial Analysis measured 59 in the Antigravity CLI. Neon scores 65.2% on Terminal-Bench 4.1. Qwen4-Max scores 48.5% on Terminal-Bench 4.0 and trails Kimi K3.1 and the US flagships on terminal coding. The two Terminal-Bench versions differ, but the gap is large either way.
Neon also leads on charts. It scores 73.5% on Chartography, against Qwen4-Max's 54.2%, which is 12 to 17 points behind the three US flagships in Alibaba's own launch chart. Google also says Neon set a new low hallucination rate for Gemini models in its internal Antigravity tests.
Qwen4-Max is strongest on long professional deliverables. It scores 21.3 on HLAB-Vals, a legal agent benchmark, nearly double Kimi K3.1's 11.7, and 46.7 on Agents' Last Exam, top of Alibaba's comparison. Neon's Vals results point the same way from a different angle, with first place on legal research and tax agent tasks. On complex documents both are capable, and Qwen4-Max's price makes it the cheaper choice for large volumes.
Caveats from early users
Neon depends heavily on the agent framework around it. On SWE-sweep it scored 12.8% with the minimal mini-SWE agent and 39.0% inside Antigravity, with the same model. Its first token can also take close to a minute at Extra High.
DataCamp's verdict on Qwen4-Max is that it's "Alibaba making a price argument more than a capability argument." It trails on terminal coding and on chart and video understanding, and "slowness and verbosity are the consistent complaints." Thinking is on by default and reasoning counts against the 262K output budget, so on complex tasks, real latency and token use are higher than the price list suggests unless you tune the thinking level.
Which to choose
Choose Gemini 4 Neon if coding is the main job: refactoring large repositories, automated testing, hard bug fixing and security code review. It's also the better fit for heavy multimodal work with complex charts, audio, video and mixed PDFs. Don't leave Extra High on for everything, since it's slow and expensive. Use Low or Medium for routine edits and save High or Extra High for architecture work and hard bugs.
Choose Qwen4-Max for document-heavy professional workflows in legal, finance and consulting: reading long documents, cross-checking many files and producing traceable reports, especially in Chinese. It also suits high-volume agent pipelines where cost per task matters most and the same context is resent often, since cached input costs $0.075 per million tokens. Turn thinking off for simple structured extraction such as JSON output, and try the Live Delivery plugin to keep long tasks on track.
If you need strong terminal coding but also want open weights or easier access from within China, Kimi K3.1 Pro is worth a look. It scores 67.2 on Terminal-Bench 4.0 in Alibaba's chart, and its weights are on Hugging Face.
Routing between models on A2Agent
Many teams split the work: a strong model for core coding, a cheaper one for long-document batch jobs, and open models for high-volume cleanup. Keeping separate accounts with several providers in different countries brings its own problems with payments, rate limits and compliance.
A2Agent gives you DeepSeek, GLM, Kimi, MiniMax, Qwen and other models behind one API key, with OpenAI-compatible and Anthropic-compatible endpoints, so you can route tasks between models without changing your code. See the Models page for current availability and prices. New accounts get $5 in free credit.
Sources
- Alibaba Cloud brings Qwen4-Max to the enterprise (Alibaba Cloud)
- Qwen4-Max: features, benchmarks, pricing and access (DataCamp)
- Alibaba's giant Qwen4-Max model trails Claude, GPT and Moonshot AI (Trending Topics)
- Google launches Gemini 4 Neon with Opus 5.5-level coding performance at 25% lower API token prices (VentureBeat)
- Google releases Gemini 4 Neon: Intelligence Index rises to third globally (Longbridge)
- Google launches Gemini 4 Neon with Carbon quality at Flash speeds (9to5Google)