
MAI Code 2 vs Qwen 4 Max: Nine Days, Two Rate Cards, One Endpoint
- OrcaNEWOrcaCyber: Hy4 Flash Abliterated / Uncensored2026-11-18$0.07 / $0.28 per 1M tokens · 180 tok/s
- tencentNEWTencent: Hy4 Flash2026-11-16$0.02 / $0.07 per 1M tokens · 223 tok/s
- GoogleNEWGoogle: Gemini 4 Carbon2026-11-10$2.00 / $10.00 per 1M tokens · 89 tok/s
- AnthropicNEWAnthropic: Claude Fable 5.22026-11-09$8.00 / $40.00 per 1M tokens · 98 tok/s
- orcaNEWOrca: OrcaC2 1.02026-11-079 tok/s
- OrcaOrcaCyber: GLM 5.3 Abliterated / Uncensored2026-11-04$3.00 / $5.00 per 1M tokens · 88 tok/s
- OrcaOrcaCyber: Hy4 Abliterated / Uncensored2026-11-04$0.80 / $1.60 per 1M tokens · 134 tok/s
- OrcaOrcaCyber: Kimi K3.1 Pro Abliterated / Uncensored2026-11-04$2.00 / $5.00 per 1M tokens · 79 tok/s
- tencentTencent: Hy42026-11-04$0.83 / $2.50 per 1M tokens · 94 tok/s
- OrcaOrcaCyber: MiniMax M3.1 Flash Abliterated / Uncensored2026-10-29$0.80 / $1.60 per 1M tokens · 118 tok/s
- OrcaOrcaCyber: GLM 5.3 Flash Abliterated / Uncensored2026-10-27$0.80 / $1.60 per 1M tokens · 68 tok/s
- OrcaOrcaCyber: Qwen3.8 Flash Next Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 94 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 · 199 tok/s
- minimaxMiniMax: M3.1 Flash2026-10-1929Intelligence
- OrcaOrca: OrcaCyber Ultra 1.02026-10-17$5.00 / $10.00 per 1M tokens · 114 tok/s
- deepseekDeepSeek: DeepSeek V4.1 Flash 10172026-10-1733Intelligence
- DeepSeekDeepSeek: DeepSeek V4.1 Flash 09102026-10-17$0.15 / $0.60 per 1M tokens · 311 tok/s
- DeepSeekDeepSeek: DeepSeek V4.1 Flash2026-10-1740Intelligence
There is a deadline hiding inside this comparison. Qwen 4 Max — which Alibaba writes Qwen4-Max, and which the API lists as qwen4-max — went metered on 16 November 2026, and the preview model strings it replaced — qwen4-max-preview and qwen4.0-max-preview — stop answering at 24:00 UTC+8 on 25 November 2026. That is nine days from today. MAI Code 2, which Microsoft released one day after Qwen4-Max went GA, at Ignite on 17 November 2026, has no such clock: it is reachable only by selecting it inside a GitHub Copilot client, at $0.50 per million input tokens and $2.50 per million output on a promotional rate that runs to 17 May 2027. So one of these two models has a migration deadline you can still miss, and the other has an expiry date you can plan around. Neither of those is a benchmark, and between them they explain most of what separates the pair.
The pricing question and the access question point in opposite directions here, and only one of them can be settled by reading a rate card.
Two models that went live a day apart and disagree about everything else
MAI Code 2 arrived at Microsoft Ignite as one of a batch of seven models on one stage, and its defining design choice is that it does not exist outside its vendor's tooling. Alibaba's Qwen4-Max left preview the day before, after a Token Plan period in Singapore, and its defining design choice is the opposite: per-token metering in six regions, four API protocols, and a documented spec sheet with the awkward parts left in.
Those twenty-four hours are the whole story of the pair. One company shipped a model optimised for a subscription it already sells. The other shipped a model optimised for an API it already operates. Every difference below follows from that, which is why the comparison is worth making at all.
The rate cards, with the discounts that expire
• Input — MAI Code 2 at $0.50 per million tokens against Qwen4-Max at $1.50 in Singapore, or $1.413 in the Global regions; MAI is a third of the Singapore rate
• Output — $2.50 against $4.50 and $4.24; the ratio narrows to 56%
• Cached input — $0.025 against $0.075 in Singapore and $0.071 elsewhere; a 3× gap, the smallest advantage in the list
• Cache writes — Microsoft publishes none, against Qwen4-Max at $1.875 per million for a five-minute write and $3.00 for an hour
• From 17 May 2027 — MAI Code 2 doubles to $1.00 / $0.05 / $5.00, at which point its input rate is still below Qwen4-Max's and its output rate is above it
• Context — 991,808 tokens of input on Qwen4-Max, dropping to 983,616 in thinking mode, against 1 million tokens on MAI Code 2 inside VS Code and Copilot CLI
• Output budget — not published by Microsoft, against a 262,144-token budget shared between reasoning and answer on Qwen4-Max
• Weights — neither is downloadable. Microsoft states this explicitly; Qwen4-Max's Max tier carries no open-weights commitment, unlike the 27B tier of the same family
The cached-input line decides long agent runs, and it cuts against the headline. Qwen4-Max's cache read is 5% of its own input rate, and its published worked example — an 8,000-token output plus a prefix read twice — turns three identical passes into $0.918 against $1.908 without caching. A team that keeps a fixed system prompt and tool schema in front of a moving question gets a far larger effective discount than the input rate implies. MAI Code 2's 95% cache discount on a lower base is comparable in kind, but Microsoft publishes no cache-write charge, which makes the two cards hard to compare on the one dimension where agentic workloads spend most of their money.
The output budget shrank on the way to general availability
The least-reported fact about Qwen4-Max's GA is that the preview gave it more room to think. At preview the shared thinking-plus-answer budget was 393,216 tokens; at GA it is 262,144. Non-thinking output is capped at 131,072. A third of the headroom disappeared at the same moment the model acquired a real price card and five new regions.
Whether that is a regression depends on your workload, but it is a documented delta that anyone still pointing a config at the preview strings needs to know about, because those strings retire in nine days. The identifier qwen4-max is the one that survives.
MAI Code 2's equivalent is not published. Microsoft documents a 1 million token context window and configurable reasoning levels — the Copilot documentation confirms reasoning depth is adjustable and notes the extended context is a VS Code and Copilot CLI feature — but there is no output ceiling in any Microsoft document, so there is nothing to compare against the 262,144 figure. That asymmetry runs through this entire matchup: Qwen4-Max tells you the awkward numbers, and MAI Code 2 tells you the flattering ones.
What the independent scores say, from the one side that has them
Artificial Analysis carries Qwen4-Max at 43 on the Intelligence Index v5.0.1, costing $3.27 per index task at 32.1 minutes per task, 55 tokens per second of output and 155.4 seconds to the first answer token, over a context window it measures at 984k. Those are mid-field numbers. The verbosity note on the same page is the more useful reading: the model generated roughly twice the output tokens of the median model on the index run, so its nominal token price understates its real cost per task.
MAI Code 2 has no Artificial Analysis row at all — no model page, so no index reading. What exists is Microsoft's launch post: four named benchmarks, three run on the company's internal MSBench platform through the GitHub Copilot CLI with subagents disabled at maximum reasoning effort, the fourth sourced from Artificial Analysis, and the results expressed as 96% of Claude Opus 5.5's performance at 5% of the cost and 98% of GPT-6.1 Sol's at 40%. No per-benchmark pass rate is published.
Line those up honestly and you cannot rank the two. Qwen4-Max's 43 on a neutral index is middling; MAI Code 2's ratios are strong on their own terms; they measure different things on different harnesses and neither vendor has run the other's. What you can say is that Qwen4-Max's number is auditable and MAI Code 2's is not, which is a statement about evidence rather than about capability.

The verbs each model was trained to perform
Microsoft's positioning for MAI Code 2 is narrow and specific: take code to production, deep debugging, multi-file implementation, repo-wide reasoning, UI work, and dashboards and visualisation. The launch post is equally specific about the edges — the models are "explicitly not yet tuned for harder problems, like fundamental science research and very long horizon agentic work" — and the blind frontend evaluation claim, 64% reviewer preference over GPT-6 Sol, is scoped to exactly that visual-coding strength.
Qwen4-Max's positioning is broader by construction, and its documentation shows it: text, image and video input, structured output, parallel tool calling, prefix completion, web search, and both explicit and implicit caching, across OpenAI Chat Completions, OpenAI Responses, Anthropic Messages and DashScope. Batch inference exists in Beijing only, and only the Batch File path carries the discount — Batch Chat is listed at the real-time rate, so the word "batch" in an endpoint name is not a price guarantee there.
That is the practical asymmetry between a coding model and a general model. MAI Code 2 is probably better at the thing it was built for, and Microsoft has internal benchmarks it will not show you. Qwen4-Max will do many more things, and Alibaba's own independent reading says it is merely competent at the coding half.
Where to run each, and what a router is actually for here
MAI Code 2 is available in GitHub Copilot across Pro, Pro+, Max, Business and Enterprise, on Windows, macOS and Linux, in Copilot CLI and Visual Studio Code. MAI Code 2 Flash is the only model in the family with an API surface — a limited-access preview in Microsoft Foundry through an OpenAI-compatible Responses API. Qwen4-Max is on Alibaba's own platform in six regions. Neither is on OrcaRouter's catalogue today, so neither claim to route them would be true.
What the catalogue does hold is Qwen3.8-Max at $2.00 and $6.00 in Singapore — the model Qwen4-Max is 25% cheaper than on both sides of the meter — plus the frontier models you would benchmark a cheap coder against: Claude Opus 5.5, Kimi K3.1 Pro, Gemini 4 Carbon. All at provider list price with 0% markup, all on one key. That matters for a specific kind of decision: if you intend to choose between MAI Code 2 and Qwen4-Max by running both on your own repositories, the comparison needs a control, and the control is measurable today while neither of the candidates is.
The routing argument that survives contact with these two models is not "put both behind one endpoint." It is that the endpoint you do control is where the fallback lives. A pipeline calling Qwen3.8-Max can fail over to a second provider without a code change; a Copilot seat cannot fail over to anything, because there is nothing underneath it to route.
How to choose without a shared benchmark
Take Qwen4-Max if the model has to answer a request your code constructs. That is the entire case and it is decisive: an API with four protocols and six regions against a model-picker entry is not a close contest for anything automated. Migrate off the preview strings before 25 November, budget for the verbosity that the index run exposed, and budget for an output ceiling a third smaller than the preview's.
Take MAI Code 2 if the developers are already in Copilot and the work is the routine enterprise repair Microsoft priced it for. At a third of Qwen4-Max's input rate and just over half its output rate, even a model needing a second attempt can win on cost per finished task — and unlike Qwen4-Max, its promotional rate has a stated end date rather than an open-ended card you do not control.
Take neither if the workload needs a score you can defend to a procurement committee. One model has a middling independent reading; the other has no independent reading whatsoever. In that situation the defensible move is to run whichever you can call against a model that is on a neutral leaderboard, on your own tasks, and compare finished-task cost rather than token price.


What would change the verdict is a single absolute number from Microsoft. A Terminal Bench 2.1 pass rate, published the way Moonshot published 65.7% on Terminal-Bench 4.0, would let a buyer place MAI Code 2 on the same axis as Qwen4-Max's 43 and stop relying on a ratio. Until then the choice is between a model you can measure and a discount you can verify, and the honest advice is to take the measurable one for anything automated and the discount for anything a person is sitting in front of.
Compared in this article1
Detected from this article · Benchmarks: Artificial Analysis · updated daily
