
MAI Code 2 vs Gemini 4 Carbon: One Discount Ends in Twelve Days, the Other Lasts Six Months
- OrcaNEWOrcaCyber: Hy4 Flash Abliterated / Uncensored2026-11-18$0.07 / $0.28 per 1M tokens · 138 tok/s
- tencentNEWTencent: Hy4 Flash2026-11-16$0.02 / $0.07 per 1M tokens · 176 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 · 97 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 · 87 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 · 121 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 · 91 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 · 195 tok/s
- minimaxMiniMax: M3.1 Flash2026-10-1929Intelligence
- OrcaOrca: OrcaCyber Ultra 1.02026-10-17$5.00 / $10.00 per 1M tokens · 118 tok/s
- deepseekDeepSeek: DeepSeek V4.1 Flash 10172026-10-1733Intelligence
- DeepSeekDeepSeek: DeepSeek V4.1 Flash2026-10-1740Intelligence
- DeepSeekDeepSeek: DeepSeek V4.1 Flash 09102026-10-17$0.15 / $0.60 per 1M tokens · 187 tok/s
On paper this is a rout. Gemini 4 Carbon sits at 60 on the Artificial Analysis Intelligence Index v5.0.1 and costs $5.48 per index task; MAI Code 2 has no independent score anywhere and costs $0.50 per million input tokens and $2.50 per million output inside GitHub Copilot. One is a measured frontier model, the other is a vendor claim with a rate card attached. But the number that decides most real decisions here is not the index gap — it is a date. Carbon's $2.00 and $10.00 rate is introductory and ends on 30 November 2026, when it doubles to $4.00 and $20.00 with nothing about the model changing. MAI Code 2's promotional rate runs until 17 May 2027. Twelve days of cheap Carbon against six months of cheap MAI Code 2 is a different comparison from the one the two spec sheets invite.
The rate cards, including what they become
• What you pay per million input tokens — Carbon $2.00 until 30 November, then $4.00; MAI Code 2 $0.50 until 17 May 2027, then $1.00
• What you pay per million output tokens — Carbon $10.00, then $20.00; MAI Code 2 $2.50, then $5.00
• Cached input — $0.025 vs an unpublished Carbon figure that Artificial Analysis computes as a 95% cache discount
• Context — 1,050,000 tokens on MAI Code 2 vs 1,048,576 on Carbon; effectively a tie
• Maximum output — 128,000 tokens on MAI Code 2 vs 65,536 on Carbon
• What it will accept — Carbon takes audio, video and documents as well as text and images; MAI Code 2's card lists text and images and nothing further
• Weights — no download on either side; Carbon is proprietary by policy, MAI Code 2 by the fact that Microsoft publishes no files

At today's numbers MAI Code 2 is a quarter of Carbon's per-token price. After 1 December it is an eighth. That is the inversion worth noticing: the gap widens precisely because Microsoft's discount outlives Google's, and any comparison written today will describe a price relationship that expires in under two weeks.
What Carbon's 60 actually measures, and what it does not
Carbon's index score is an average of nine evaluations run by Artificial Analysis, including Terminal-Bench 4.0, FrontierCode-AA 2.0, AA-Briefcase v1.1, AutomationBench-AA and Terminal-Bench-Science 0.1. It is the third-highest score on that board and it is independently produced. The supporting numbers are less flattering than the rank: 34.2 minutes per index task, 74.42 seconds to first answer token, 60 tokens per second of output, and $5.48 per task. Carbon is a slow, deliberate model that costs real money to run to completion.
That combination — top-three score, bottom-half throughput — is what a multimodal model looks like when it is doing the work rather than summarising it. It is also the reason a per-token comparison flatters Carbon: at 60 tokens per second, a long agentic task is minutes of wall clock whether or not anyone is watching.

Where Carbon is genuinely ahead of anything Microsoft is claiming is breadth. The model accepts text, images, audio, video and documents and returns text, with a 1M-token context and native tool use. Microsoft's model card lists text and images in, text out. If the workload involves watching a screen recording of a failing test, or reading a scanned specification alongside a codebase, Carbon is the only one of the two that claims the input at all.
What MAI Code 2's numbers are and are not
Microsoft publishes three pass rates at maximum reasoning effort: 73.2% on DeepSWE 1.1, 84.6% on TerminalBench 2.1 and 43.1% on SWE Atlas. All three were produced on MSBench, Microsoft's internal evaluation platform, through the GitHub Copilot CLI at version 1.0.102 with subagents disabled. There is no fourth-party confirmation and no Artificial Analysis row. The model card is candid about the scope too: optimised for English code production only at release, not tuned for very long-horizon agentic work, and not intended for unsupervised consequential actions.
The architecture is at least specific enough to reason about. MAI Code 2 is a sparse mixture-of-experts transformer with about 2.8 trillion total parameters and 77 billion active, plus a 3-billion embedded vision encoder, pre-trained on roughly 50 trillion tokens with long-context extension to a million. That is a large model by any measure and a cheap one to call, which is the entire pitch.
The distribution gap is the real one
Gemini 4 Carbon has been generally available on the paid Gemini API since 10 November 2026, after twelve days of gated access through Google's Fairwind Program. It has a stable model ID, a rate card, an AI Studio presence and enterprise availability.

MAI Code 2 has GitHub Copilot. That is not a small thing — Copilot Pro, Pro+, Max, Business and Enterprise, across VS Code, Visual Studio, the CLI, the cloud agent, JetBrains, Xcode and Eclipse — but it is one vendor's surface. MAI Code 2 Flash is in limited-access preview in Microsoft Foundry through an OpenAI-compatible Responses API. MAI Code 2 itself has no Foundry deployment, no published endpoint and no date for one. Microsoft's announcement says further availability is coming and names no platforms.
So the practical answer to "which should I use" depends on a question that has nothing to do with benchmarks: are you building inside Copilot, or are you building an application that calls an API? In the second case there is only one model in this comparison you can actually call today.
Where the two models are each the right answer
• Everyday enterprise repair inside a Copilot workflow — Carbon is not in Copilot and MAI Code 2 is, at a quarter of Carbon's price and an eighth of it from December. This is not close.
• Multimodal or media-heavy input — Carbon wins by default, because MAI Code 2 does not claim audio, video or document input.
• Long-horizon autonomous agent work — neither is the obvious pick. Carbon is slow and expensive per task at 34 minutes and $5.48; Microsoft says outright that MAI Code 2 is not tuned for it.
• Cost-sensitive batch work at scale — the price doubles on Carbon in twelve days and on MAI Code 2 in six months. If the workload is steady-state, that asymmetry favours the MAI side; if it must be callable from your own code, it does not.
• Workloads needing a full independent record — Carbon has one, MAI Code 2 has none. A procurement review that requires third-party evidence can only clear one of these two models.
Pricing both sides on one key
Gemini 4 Carbon is on OrcaRouter at $2.00 input and $10.00 output per million tokens, provider list price passed through at 0% markup — which means the 1 December repricing to $4.00 and $20.00 lands on your bill the day Google changes it, with no reseller lag and no need to renegotiate anything. MAI Code 2 is not on our catalogue, and the only published route to it is GitHub Copilot.
That asymmetry is exactly the case for routing rather than committing. Carbon's price is about to double; MAI Code 2's is not about to be callable from your own stack. Sitting behind one endpoint, the Carbon line item can be swapped or split the moment the new rate card takes effect, and a second model can carry the traffic that used to justify the cheap rate — no code change, no second contract. Automatic failover matters here for a second reason: Carbon at 74 seconds to first token and 34 minutes per task is a long-running dependency, and a run that stalls is a run you want somewhere else.
What to do before 30 November: work out what fraction of your Carbon traffic is routine enough that a model without an independent score could plausibly hold it, and what fraction genuinely needs the multimodal input or the 60-point index. Then price the second fraction at $4.00 and $20.00, because that is what it will cost. The first fraction is where MAI Code 2 is worth a look — inside Copilot, today, at $0.50 and $2.50.
