
MAI-Code-2 Is Already in Copilot's Model Catalog. Microsoft Hasn't Said a Word.
- GoogleNEWGoogle: Gemini 4 Carbon2026-11-10$2.00 / $10.00 per 1M tokens · 90 tok/s
- AnthropicNEWAnthropic: Claude Fable 5.22026-11-09$8.00 / $40.00 per 1M tokens · 97 tok/s
- orcaNEWOrca: OrcaC2 1.02026-11-0711 tok/s
- OrcaNEWOrcaCyber: GLM 5.3 Abliterated / Uncensored2026-11-04$3.00 / $5.00 per 1M tokens · 88 tok/s
- OrcaNEWOrcaCyber: Hy4 Abliterated / Uncensored2026-11-04$0.80 / $1.60 per 1M tokens · 134 tok/s
- OrcaNEWOrcaCyber: Kimi K3.1 Pro Abliterated / Uncensored2026-11-04$4.00 / $20.00 per 1M tokens · 86 tok/s
- OrcaOrcaCyber: MiniMax M3.1 Flash Abliterated / Uncensored2026-10-29$0.80 / $1.60 per 1M tokens · 125 tok/s
- OrcaOrcaCyber: GLM 5.3 Flash Abliterated / Uncensored2026-10-27$0.80 / $1.60 per 1M tokens · 67 tok/s
- OrcaOrcaCyber: Qwen3.8 Flash Next Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 92 tok/s
- OrcaOrcaCyber: MiMo V2.6 Flash Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 43 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 · 192 tok/s
- minimaxMiniMax: M3.1 Flash2026-10-1929Intelligence
- OrcaOrca: OrcaCyber Ultra 1.02026-10-17$5.00 / $10.00 per 1M tokens · 134 tok/s
- deepseekDeepSeek: DeepSeek V4.1 Flash 10172026-10-1733Intelligence
- DeepSeekDeepSeek: DeepSeek V4.1 Flash2026-10-1740Intelligence
- OrcaOrca: OrcaCyber Zero 1.52026-10-10$3.00 / $7.50 per 1M tokens · 129 tok/s
- openaiOpenAI: GPT-6.1 Sol2026-09-2955Intelligence
- anthropicAnthropic: Claude Sonnet 5.52026-09-2855Intelligence
- typesafeTypeSafe: Jev 1.132026-09-24$0.04 / $0.00 per 1M tokens · 187 tok/s
The firmest evidence that MAI-Code-2 and MAI-Code-2-Flash exist is an error message. On 14 November 2026 a maintainer of Vekil, an open-source gateway that proxies GitHub Copilot, filed a bug: Copilot's /models endpoint lists both mai-code-2 and mai-code-2-flash, advertising both /chat/completions and /responses — but every /chat/completions call to them returns 400 no_available_model_endpoints while /responses succeeds. Three days later a second PR in the same repo recorded that neither model lists ws:/responses, so a native websocket connection dies on the first turn. Both of those are the fingerprints of a model that is wired into a production catalogue without being finished, and neither is an announcement. Meanwhile microsoft.ai still presents MAI-Code-1.1-Flash as the newest coding model Microsoft makes. This is what can be dated, what it implies, and what Microsoft has conspicuously not said — and nothing below should be read as a release.
Four receipts, all public, none of them a blog post
Microsoft has published nothing about MAI-Code-2. So the case has to be built out of artifacts that were never meant to be read as a launch, which is exactly why they are worth reading. In date order:
• 4 November 2026 — mai-code-2-flash first appears in the automated model-inventory diff committed to github/gh-aw, GitHub's own agentic-workflows repository. It enters with all four cost fields set to zero, which is what a newly discovered catalogue entry looks like before anyone prices it.
• 7 November 2026 — the same inventory job, three days later, carries real numbers: input 1e-07, output 5e-07, cache-read 1e-08, cache-write 0, with wire_api: responses. That is $0.10 per million input tokens, $0.50 per million output, $0.01 per million cached reads. A zero-to-priced transition inside one week is a pricing sheet being filled in, not a rumour.
• 13 November 2026 — a documentation PR in microsoft-foundry/fine-tuning adds MAI-Code-2-Flash alongside MAI-Code-1.1-Flash in the managed reinforcement-fine-tuning track, and prints the two identifiers verbatim: training ID mai-code-2-flash-2026-11-17 and catalogue identifier Microsoft/mai-code-2-flash/2026-11-17. The PR is careful to say the date "identifies a model version, not a launch date." It is still the only place a build stamp for this model exists in public.
• 15 November 2026 — Azure Developer CLI merges a catalogue change that replaces MAI/MAI-Code-1.1-Flash with MAI/MAI-Code-2-Flash in the published RLE model catalogue, then ships it the next day as a forced update (extension 0.8.54-preview). The compatibility note is blunt: the retired identifier is now rejected by catalogue validation, and existing configurations "must explicitly select a supported model," with no migration path offered. That is a vendor-adjacent toolchain retiring the old model name three days before the new one is announced.
Add the trigger for this piece: on 17 November the TestingCatalog account on X put "MAI-Code-2 and MAI-Code-2-Flash, today!" at the top of its daily model forecast. That account has a track record on Microsoft releases, and its forecast is the reason to write this now rather than after the fact. It is also still a forecast. Treat it as the weakest receipt in the stack, not the strongest.

What the artifacts say about the model itself
None of the four receipts contains a benchmark, a context window, a parameter count, or a licence. What they do establish is the shape of the thing, and the shape is more informative than it sounds.
There are two variants, not one. A full MAI-Code-2 and a MAI-Code-2-Flash, listed side by side and handled side by side by every downstream tool that touched them. Microsoft's coding line has never shipped a paired standard/Flash release — MAI-Code-1-Flash and MAI-Code-1.1-Flash were single models — so a two-tier launch is a genuine change in how the family is packaged. The Vekil reports put both under catalogue vendor "Experimental", while MAI-Code-1.1-Flash sits under "Microsoft". That distinction matters if you administer Copilot model policies: an Experimental vendor entry is not the same object as a GA Microsoft one, and policy allow-lists keyed to GA models will not cover it by default.
The pricing signal is the concrete part. At $0.10 input and $0.50 output per million, the Flash variant would land at exactly half the input price and just over 41% of the output price of MAI-Code-1.1-Flash — 58% cheaper on output, half price on input — which GitHub's own Copilot pricing table lists at $0.20 in, $0.02 cached, $1.20 out. Run it through a realistic agent turn — 200K input tokens of repository context, 10K tokens of diff and explanation — and 1.1 Flash costs about $0.052 while the Code 2 Flash numbers come to about $0.025. On a team doing a thousand such turns a day, that is roughly $25 a day against $52. Those are arithmetic on catalogue constants, not a quote: no Microsoft price page lists either model, and a query of the Azure retail price API today returns zero meters matching "Code 2". If Foundry does publish a list price, expect it to differ from what a Copilot cost map carries, because the two are billed differently.
Then there is the endpoint detail, which is the least glamorous and most decision-relevant item in the whole stack. Both models advertise /chat/completions and /responses, serve /responses, and fail on /chat/completions with no_available_model_endpoints. Whatever ships, it is a Responses-API model first. Any harness, proxy, or internal gateway your team runs that assumes Copilot models are Chat-Completions-native will need the same fix two open-source projects had to write this week — route Microsoft and Experimental catalogue entries that advertise /responses straight to /responses, and give up native-Chat-only extensions such as Anthropic-style cache breakpoints on those routes.

The predecessor sets the bar, and the bar is real
MAI-Code-2 does not arrive into an empty category, so the useful comparison is against the model it replaces. MAI-Code-1-Flash launched on 2 June 2026 at Microsoft Build as a lightweight agentic coder for GitHub Copilot. MAI-Code-1.1-Flash followed on 11 August 2026, claiming 25% better token efficiency, a 22% gain on Terminal-Bench 2.1 in Copilot CLI, and a 15% improvement on .NET work. On Microsoft's own model page, 1.1 Flash reports 72.6% on SWE-Bench Verified at 8.6K tokens per solution, 62.9% on Terminal-Bench 2.1, 74.1% on Text2WebApp, and 42.1% on Screenshot2WebApp — all vendor-reported, none independently reproduced, and Microsoft's page itself concedes the model trails Claude Haiku 4.5 on Vision2Web (11.5% against 13.7%). The same page describes a 256K context window and an on-device variant at 3-bit quantization that keeps 70.8% SWE-Bench Verified, with Microsoft recommending machines with more than 120GB of RAM to run it locally.
That is the bar for a "2": roughly 73% on SWE-Bench Verified, mid-60s on Terminal-Bench, strong on mockup-to-app generation, weak on vision-heavy web tasks, and cheap enough that token efficiency is the headline instead of raw capability. A successor that only matches those numbers while halving the price would still be a commercially significant release, because the price is the thing agent loops actually consume. A successor that beats them is a different story. Nothing public yet says which one this is.
What Microsoft has not said, and why the silence is the story
Four artifacts point one way; the vendor's own surfaces point the other. As of today, microsoft.ai's model sitemap lists exactly one coding model — the MAI-Code-1.1-Flash page — and its news archive contains no post about a second generation. GitHub's Copilot documentation lists MAI-Code-1.1-Flash as GA and, in its retirement table, records MAI-Code-1-Flash as retired on 10 September 2026 with "Latest MAI-Code model" as the suggested alternative. The GitHub Copilot changelog, which did carry the November model additions for Claude Fable 5.2 and Gemini 4 Carbon, has no MAI-Code entry. Artificial Analysis, which tracks most frontier coding models, carries no MAI-Code row at all.

There is one more signal pointing the same direction, and it is worth flagging as weaker than the rest. The LiteLLM price map carries a deprecation date of 2026-12-09 on the Foundry-hosted MAI-Code-1.1-Flash entry. That is community-maintained metadata, not a Microsoft lifecycle notice — Microsoft's own model lifecycle page does not list MAI-Code models at all — but it is a third party independently concluding that 1.1 Flash has a shelf life measured in weeks.
The charitable reading of all this is mundane and probably right: Microsoft is staging a launch the way it staged 1.1 Flash, wiring catalogues and downstream tools first and publishing the post second. The uncharitable reading is that some of this is unfinished plumbing that leaked early — the 400 on /chat/completions is not a design choice anyone would make deliberately. Either way, the practical advice is identical: nothing here is safe to build a production dependency on today, and everything here is worth knowing before the announcement lands, because the announcement will arrive with a price and the catalogue already tells you roughly what it is.
If you administer Copilot, one thing needs doing now
Independent of whether MAI-Code-2 ships today or in three weeks, the Azure Developer CLI change is already live and already breaking. Extension 0.8.54-preview was published as a forced update on 16 November 2026, and it removes MAI/MAI-Code-1.1-Flash from the shared model catalogue in favour of MAI/MAI-Code-2-Flash. Existing configurations that name the old identifier now fail catalogue validation, and there is no manifest rewrite or session migration. If your team pins that identifier in rle.toml, in a CI job, or in a saved Loom session, the fix is a manual edit and it is due now, not at announcement time. Because the catalogue is downloaded on every CLI invocation rather than shipped inside the binary, the change reached you without a version upgrade you had to accept.
The Copilot side is quieter. If the models are visible in your tenant under an Experimental vendor label, an administrator still has to enable them at the policy level before anyone can select them in VS Code or Copilot CLI — the same gate that applied to MAI-Code-1.1-Flash when it went GA. Until that policy exists, the catalogue entry is visible to tooling and useless to developers, which is a fair description of where this sits right now.
Where it will be callable, and the honest answer about today
Availability, when it comes, will almost certainly follow the pattern of both predecessors: GitHub Copilot first — VS Code, Copilot CLI, the Copilot app — with Microsoft Foundry following for teams that want a direct API, the same split that put MAI-Code-1.1-Flash in Copilot in August and on Foundry's price sheet in October. That is inference from two prior releases and it is labelled as such. No Microsoft page currently commits to any of it.
For readers who route coding models across providers rather than standardising on one vendor's assistant, the workable position is the boring one: keep the comparison set on one key and swap the unproven model in behind a test, not in front of production. OrcaRouter carries the coding models this launch will actually be judged against — GPT-5.3-Codex, GPT-5.2-Codex, Kimi K2.7 Code — at list price with no markup on top, which means a vendor price change shows up in your bill the day it happens rather than at the next contract renewal. It does not carry MAI-Code-2 or MAI-Code-2-Flash, and it does not carry MAI-Code-1.1-Flash either: those are Copilot- and Foundry-only models today, and no third-party router has them. Anything else you read claiming otherwise is describing a proxy of a Copilot subscription, which is a different product with different terms.
What we would want before recommending anyone move: a Microsoft post with numbers, a Foundry price meter, and an independent score. Two of those three can be checked in a browser in under a minute, which is the useful thing about this particular leak — it is unusually falsifiable. When the announcement lands, either the catalogue numbers above hold or they don't, and either the benchmark claims clear 72.6% on SWE-Bench Verified or they don't. Until then, the honest summary is the one the error message gave us: the model answers on /responses, it is priced at ten cents, and Microsoft has not yet decided to tell anyone.
Compared in this article1
Detected from this article
