MAI Code 2 and Copilot Fusion: Microsoft Ignite 2026 coding models, Terminal-Bench claim
Microsoft's Ignite keynote made MAI Code 2 and Code 2 Flash generally available in GitHub Copilot; here is what shipped, who gets it and which claims are only Microsoft's.
· 6 min read
Written by AI from the sources linked below. How we use AI
On this page
Microsoft used its Ignite keynote on 17 November 2026 to pitch Copilot as "the new operating system for work", and the developer-facing part landed the same day: two new in-house coding models, MAI Code 2 and MAI Code 2 Flash, are generally available in GitHub Copilot, and GitHub says its Copilot Fusion router now sits at 71.2% on a public Terminal-Bench 4.1 leaderboard. The pitch arrives a week after a large Microsoft 365 outage, which CEO Satya Nadella addressed at the start of the talk.
Key points
- MAI Code 2 and MAI Code 2 Flash became generally available on 17 November in GitHub Copilot and in Microsoft Foundry.
- GitHub says Copilot Fusion Max reached a 71.2% resolution rate on Terminal-Bench 4.1, 5.4 percentage points ahead of Opus 5.5 and Fable 5.2, at 57% lower cost than Opus 5.5 and 88% lower than Fable 5.2 (GitHub's figures).
- Microsoft says Agent 365 now holds more than 50 million agents, and that its Perception agentic security system and Cobalt 200 VMs are generally available.
- Microsoft announced the Maia 300 chip, claiming 40 percent more tokens per watt than leading GPU-based systems; no availability date was given.
- Pricing and many availability dates were not stated in the keynote post.
What happened
Nadella's keynote post on the official Microsoft blog describes Copilot as one experience that brings together Chat, Cowork, Code and Autopilot, with a Home screen as the starting point and Copilot Workspaces as a shared, multiplayer space where people and agents work from the same context. Role-specific extensions cover sales, legal, finance and service, there is a Dynamics 365 plugin, and a Salesforce tie-up that Microsoft calls Copilotforce. The Register adds that Nadella said apps are becoming plugins to Copilot and that "all of Office is built into it".
Underneath it is Microsoft IQ, which Microsoft describes as a platform that exposes Copilot's context system to organisations, drawing on Fabric, Foundry, Microsoft 365 and third-party data, with "hundreds" of new connectors (no exact count given). Work IQ, covering an organisation's people, documents, email, meetings and projects, is the data layer Microsoft pitches as an organisation's most important database.
The coding models
For developers, the concrete news is in the GitHub changelog. MAI Code 2 is described as a general-purpose coding model for deep debugging, multi-file implementation, refactoring and repository-wide reasoning. MAI Code 2 Flash is the cheaper, lighter model for explanations, targeted edits and quick iteration.
- MAI Code 2: Copilot Pro, Pro+, Max, Business and Enterprise.
- MAI Code 2 Flash: all of those plus Copilot Free and Student.
- Where: the model picker in VS Code, Visual Studio, the Copilot CLI, the Copilot cloud agent, the Copilot app, github.com, GitHub Mobile, JetBrains, Xcode and Eclipse.
- Billing: usage-based, with rates left to GitHub's "Models and pricing for GitHub Copilot" documentation; the changelog lists no multipliers.
- Admins: Business and Enterprise administrators control access through the model policy. New models are switched on automatically unless the global default is disabled or the model is explicitly disabled.
Microsoft's keynote post says Code 2 is priced at a "fraction of frontier cost" without a number, and that Code 2 and Flash are also generally available in Foundry.
Copilot Fusion and the benchmark claim
GitHub's second post covers Copilot Fusion, which chooses and coordinates models per task while the user sets an effort level. Fusion now appears on the public Terminal-Bench 4.1 leaderboards hosted by Stanford, Harbor and the Laude Institute. GitHub says Fusion Max reached 71.2%, 5.4 percentage points ahead of Opus 5.5 and Fable 5.2, for 88% less than Fable 5.2 and 57% less than Opus 5.5. The Register summarises the claim as ranking above Opus 5.5 at under half the cost. Those are GitHub's numbers; we did not find an independent run of the same setup.
New in the mix: Claude Sonnet 5.5 in Fusion Max, GPT-6.1 Terra in Fusion High, and MAI-Code-2 in Fusion Standard. Fusion can now be picked in the model picker on github.com, in Visual Studio, and in the iOS and Android GitHub Mobile beta apps, with production mobile apps promised shortly.
Background
Microsoft has been shipping its own MAI models through GitHub Copilot since earlier this year. GitHub Copilot has simultaneously been pruning other models: we covered the six models Copilot retires on 14 December in an earlier piece. At Ignite, Microsoft said Foundry will offer GPT-6.2, and OpenAI the same day published the details we cover in GPT-6.2 Astra and Sol: prices, context window and independent tests.
The keynote also opened with trust. The Register reports that Nadella apologised for the Microsoft 365 outage the week before and framed Ignite around keeping humans at the centre, an organisation owning its own intelligence, and trust. We covered the outage week in US outage week: what each cause is.
How it works
Fusion is a routing layer. Instead of choosing one model, you choose an effort tier: Standard, High or Max. GitHub names MAI-Code-2 in Standard, GPT-6.1 Terra in High and Claude Sonnet 5.5 in Max, and Copilot decides which model handles each step. That is why a benchmark score belongs to the whole system, not to any single model, and why cost comparisons depend on how many calls the router makes.
The rest of the keynote
- Agent 365: a control plane for governance, identity and security of enterprise agents, with more than 50 million agents, and a Business Analytics view for tracking AI spend (no date or price given).
- Perception: an agentic security system with red, blue and green agent teams running Playbooks, now generally available. Cyber 2 Flash, a security-focused model, is being integrated with it.
- Decisions 1.1: generally available, with more than 4,000 customers having used it, per Microsoft.
- Silicon: Cobalt 200 VMs are generally available. Maia 300 was announced with a 40 percent more tokens per watt claim and no availability information.
- AI Transparency Initiative: collaboration in NIST's agent standards work and an effort with 40 organisations on standards.
What it means
For developers: there is now a Microsoft-built model in the same picker as the frontier labs' models, and a router that mixes them. The trade-off is that you give up knowing which model produced a change. Test MAI Code 2 Flash on your own repository before turning it on for a team.
For admins: because new models are enabled by default unless you disabled the default, check your policy now, especially if you restrict models for data or licensing reasons.
For Microsoft's competitors: the Register notes that emphasis on customer ownership of AI may sit awkwardly with a company that has big cloud commitments to run AI workloads, and that frontier labs may not like it. It also raises the open question of who owns AI-generated code.
What we don't know yet
- Pricing: neither post gives per-unit rates for MAI Code 2, and the keynote gives no prices for Copilot, Agent 365 or Maia 300.
- Whether Fusion's leaderboard result holds up in independent testing.
- Availability dates for Copilot Workspaces, Work IQ, Maia 300 and Microsoft IQ connectors.
- The Ars Technica report on the keynote would not open for us, so we did not use it.
What to watch next
- GitHub's models and pricing documentation for MAI Code 2 rates.
- Independent Terminal-Bench 4.1 runs of Fusion and the MAI models.
- Production GitHub Mobile support for Fusion.
- Microsoft's follow-up on the Microsoft 365 reliability promise.
Sources
Written by
The official account of the people who run audiotale.net: announcements, conversation starters and the newsroom's drafts. Posts written with AI say so. Every other account here is a real person.
Read next
Microsoft Voluntary Exit Offer: about 60,000 eligible in US, Canada, India and UK
Microsoft's new time-limited buyout covers staff at level 70 and below, roughly a quarter of its workforce according to Bloomberg, after a smaller July round of role cuts.
audiotale.net team6 min read

Microsoft-Decision-1: 9B Qwen-based model scores answer options; 35x faster than GPT-6 Sol, Microsoft says
Microsoft's new Foundry and OpenRouter model returns calibrated probabilities for fixed answer options instead of text; the speed and accuracy claims are the company's own.
audiotale.net team6 min read
Microsoft blames China-linked Storm-0940 for ConnectWise Cloud Backup breach, says attackers used ARTEX AI framework
Microsoft says Storm-0940 used the open-source ARTEX framework against hundreds of organisations after stealing ConnectWise app credentials; what to check in your Microsoft 365 tenant.
audiotale.net team8 min read