生成的状态卡将 MAI Code 2 与 GLM-5.5 进行对比:MAI Code 2 自 2026 年 11 月 17 日起推出,输入价格为每百万 token 0.50 美元,输出价格为每百万 token 2.50 美元;GLM-5.5 没有模型卡、API 标识符或价格,并被展示在 Z.ai 实际销售的模型 GLM-5.3 旁边。
Guides & Insights

MAI Code 2 对比 GLM-5.5:这两款模型中的一款提供了费率卡

作者

jinhao.song

发布日期

最新模型 · 20查看全部模型 →
基准测试:Artificial Analysis · 每日更新
返回全部文章

MAI Code 2 shipped on 17 November 2026 with a 24-page model card, a published promotional rate of $0.50 per million input tokens and $2.50 per million output, and a deployment in GitHub Copilot. GLM-5.5 has none of those things. Z.ai's own API platform tops out at GLM-5.3 and GLM-5.3-Flash; the lab's Hugging Face organisation has no checkpoint carrying the 5.5 number; there is no rate card row, no API identifier and no model card. So the matchup the title promises cannot be run, and pretending otherwise would mean inventing a specification. What can be run is MAI Code 2 against GLM-5.3, the model Z.ai actually sells — plus a precise account of what would have to appear before "GLM-5.5" became something you could call, and why a reader arriving on this page should care about the difference between a roadmap and a product.

What exists on each side, checked this week

• MAI Code 2 — released 17 November 2026, model card published, weights not distributed, API access inside GitHub Copilot only

• GLM-5.5 — no release, no model card, no identifier, no price, no weights, no vendor page using the name

• GLM-5.3 — released August 2026, open weights on Hugging Face under Z.ai's own organisation, API live on Z.ai's platform

• GLM-5.3-Flash — released 26 August 2026, the high-volume sibling, also open weights

The absence is not a technicality. A model exists when a lab will take your money or hand you the weights. Z.ai does both for GLM-5.3 and neither for anything numbered 5.5. Every claim circulating about GLM-5.5 sits on a release-window forecast rather than a vendor statement, and windows have slipped before in this line — the August window that produced GLM-5.3 was widely expected to produce a model called GLM-5.4.

The comparison you can actually run: MAI Code 2 vs GLM-5.3

Two-column card contrasting MAI Code 2, released 17 November 2026 at $0.50 and $2.50 per million tokens with 1,050,000-token context and no distributed weights, against GLM-5.3, released August 2026 with open weights, a $1.26 and $3.96 rate row and an Artificial Analysis Intelligence Index of 32.

These two are closer in spirit than in price. Both are aimed at repository-scale software work. Both are closed in MAI Code 2's case and open in GLM-5.3's, which is the sharpest structural difference between them.

• Input price — MAI Code 2 at $0.50 per million tokens (promotional, until 17 May 2027) vs GLM-5.3 at $1.26 according to OrcaRouter's structured rate row, $1.40 according to the vendor's own descriptive text

• Output price — $2.50 vs $3.96 by the structured row, $4.40 by the descriptive text

• After the promotional window — MAI Code 2 moves to $1.00 and $5.00 on 17 May 2027; GLM-5.3 has no announced change

• Context — GLM-5.3 takes a 1,000,000-token window; MAI Code 2 takes 1,050,000, so the extra 50,000 is the only place the Microsoft model is ahead on this line

• 最大输出 — 两者均为 128,000 个 token

• What it will accept — GLM-5.3 is text in, text out; MAI Code 2 also takes images in, which matters only if your tickets carry screenshots

• Weights — MAI Code 2 distributes none; GLM-5.3 is downloadable, so you can run it yourself

• Independent score — MAI Code 2 has none; GLM-5.3 posts 32 on the Artificial Analysis Intelligence Index v5.0.1 at $5.01 per index task

The price discrepancy on the GLM side is worth flagging rather than smoothing over: the catalogue's structured figure and the vendor page's prose disagree, and which one you pay depends on how the rate card is applied. That is a reason to check the number at the moment you commit, not a reason to pick one and call it settled.

What GLM-5.3's independent score tells you that MAI Code 2's vendor numbers do not

Screenshot of the Artificial Analysis page for GLM-5.3 showing an Intelligence Index of 32, an August 2026 release, open weights and a $1.40 per million input token rate.

GLM-5.3's 32 is not impressive on its own. It sits below Claude Opus 5.5 at 59, Gemini 4 Carbon at 60 and GPT-6.1 Sol at 55, and it is 22 minutes and $5.01 per index task to get there. But it is a number produced by someone with no stake in Z.ai's marketing, on the same nine-evaluation basis as everyone else's, and that makes it comparable to the rest of the board in a way Microsoft's 96%-of-Opus claim is not.

Microsoft's figures — 73.2% on DeepSWE 1.1, 84.6% on TerminalBench 2.1, 43.1% on SWE Atlas, all at max reasoning effort — came out of MSBench, Microsoft's internal harness, through the GitHub Copilot CLI. The model card is unusually forthcoming about method and unusually silent about independent confirmation. When one side of a comparison has a third-party score and the other has a first-party pass rate, the honest reading is that you know where GLM-5.3 sits and you know what Microsoft says about MAI Code 2.

The open-weights question decides more than the score does

Screenshot of the Hugging Face model card for zai-org/GLM-5.3, showing an openly published repository under Z.ai's own organisation with a TextGeneration pipeline tag, and the model card's statement that GLM-5.3 uses the same base model as GLM-5.2 and takes its gains from post-training.

This is the difference that actually changes architecture rather than procurement. GLM-5.3 is downloadable. You can put it on your own hardware, fine-tune it on your own repository history, keep the traffic inside your network, and stop paying per token at the point where your volume exceeds the cost of a GPU. Z.ai also ships a customisation path — Microsoft's equivalent is MAI Licensed Weights, announced the same day as MAI Code 2, which lets an organisation run a customised MAI model inside its own Azure environment. The two are not the same offer. One is a file you can download and run wherever you like; the other is a commercial arrangement that keeps the model inside a Microsoft cloud and keeps the weights in Microsoft's hands.

For a team whose constraint is data residency, GLM-5.3 answers the question outright. For a team whose constraint is that it already runs on Azure and wants a supported model with a vendor behind it, MAI Licensed Weights is the closer fit — and it is a separate purchase from the Copilot subscription that gets you MAI Code 2 by the token.

Where MAI Code 2 is and is not available

On the Copilot side the surface is wide but the door is narrow: five subscription tiers, from Pro up to Enterprise, and every editor GitHub ships an extension for. What is missing is the thing an application needs — a model ID and a URL. MAI Code 2 Flash has a limited-access Foundry preview on an OpenAI-compatible Responses API; the large model has no Foundry deployment, no published endpoint and no date for one, and Microsoft says only that broader availability is coming.

GLM-5.3, by contrast, is on Z.ai's own API and available as weights, which means a dozen hosting arrangements exist that neither vendor controls. That is the practical difference between a model with an ecosystem and a model with a launch.

If GLM-5.5 does appear, what would change

Three things would have to be true before it became a real comparison. A model card or a vendor announcement, because a name in a leak thread is not a release. A rate card, because the only decision a buyer makes is per-token cost against per-task cost. And weights or an endpoint, because a model you cannot call is a press release. The likely shape, if Z.ai follows its own pattern, is a GLM-5.3-class text model with a higher index score and open weights — which would make it a genuine alternative to MAI Code 2 for anyone who needs to run the model rather than rent it, and no alternative at all for anyone working inside Copilot.

Running both sides on one key

GLM-5.3 is on OrcaRouter's catalogue at the pass-through rate, with 0% markup, alongside Claude Opus 5.5, Gemini 4 Carbon and GPT-6.1 Sol — so the open-weights side of this comparison can be measured against the closed frontier on a single key without a second contract. That is the useful capability in a matchup where one model is a promise: you can price the alternatives you actually have, today, and see what the model that does not exist would have to beat. MAI Code 2 is not on our catalogue; the published route is GitHub Copilot.

For a reader who came here to choose: if you want a model that exists, has open weights and a published price, GLM-5.3 is the answer on the Z.ai side. If you want the cheapest per-token coding model with a Microsoft support path and you work in Copilot, MAI Code 2 is the one shipping today. GLM-5.5 is neither, and the honest thing to do with a name that has no model behind it is to wait for the model card rather than plan around the name.

本文中的对比1

根据本文内容识别 · 基准测试:Artificial Analysis · 每日更新