Models

Microsoft AI Launches MAI Code 2 and Flash on Cost-per-Task Math

Microsoft AI's MAI Code 2 claims 98% of GPT-6.1 Sol's performance at 40% of the cost, with a Flash variant beating comparators on both pass rate and price per task.

Microsoft AI Releases MAI Code 2 and MAI Code 2 Flash for Enterprises
Microsoft AI Releases MAI Code 2 and MAI Code 2 Flash for EnterprisesStockSnap / pixabay
By James Calloway6 min read

Why it matters

  • Microsoft announced MAI Code 2 and MAI Code 2 Flash on November 17, 2026, claiming 98% of GPT-6.1 Sol's performance at 40% of its cost.
  • MAI Code 2 averaged 63.2% pass rate at $0.309 per task across four benchmarks; Claude Opus 5.5 averaged 65.3% at $6.306.
  • MAI Code 2 Flash reported 61.0% pass rate at $0.052 per task, beating GPT-6 Luna and Claude Haiku 5.5 on both metrics.
  • Microsoft cost figures use a 50% promotional price, and benchmarks ran on its internal MSBench platform.
  • A Flash derivative, MAI Work Flash, scored 0.77 in Power BI DAX query-generation evaluations versus 0.54 for GPT-6 Astra.

Microsoft AI announced two enterprise coding models, MAI Code 2 and MAI Code 2 Flash, on November 17, 2026, claiming MAI Code 2 delivers 98% of GPT-6.1 Sol's performance at 40% of its cost and 96% of Claude Opus 5.5's performance at 5% of its cost across four benchmarks.

The company is not claiming benchmark supremacy. It is claiming price-performance. That distinction defines the entire release, and Microsoft is unusually explicit about where the argument holds and where it does not.

Microsoft positions the models for "everyday enterprise coding" rather than frontier research. MAI Code 2 targets production changes, deep debugging, multi-file implementation, repository-wide reasoning and user interfaces. The Flash variant handles high-volume work: explaining code, making targeted edits, applying routine fixes and iterating quickly. The split separates deep implementation from frequent tasks where per-task cost compounds at scale.

Microsoft is also explicit about the boundary. Neither model has been tuned for fundamental science research or very long-horizon agentic work. That limitation, stated by the company itself, defines the release's scope around the coding work customers do daily — and quietly excludes the hardest use cases where rivals might still lead. In plain words: Microsoft is competing on routine work at cut-rate prices, not on the frontier.

What do the benchmark numbers actually say?

Across DeepSWE 1.1, Terminal-Bench 2.1, SWE Atlas and FrontierCode v1.1, Microsoft reports a 63.2% average pass rate and $0.309 average cost per task for MAI Code 2. GPT-6.1 Sol averaged 64.1% at $0.761 per task, while Claude Opus 5.5 averaged 65.3% at $6.306.

The pass-rate gaps are small. The cost gaps are not. Claude Opus 5.5 costs roughly twenty times more per task than MAI Code 2 while scoring roughly two points higher on average. That arithmetic is the release's core sales pitch.

But the comparisons come with conditions Microsoft itself discloses. All models in the coding comparisons ran at maximum reasoning effort. The cost figures for Microsoft's models use a 50% promotional price. And the company cautions that changing the workload, model configuration, reasoning effort or pricing assumptions can change the comparisons.

That caution deserves weight. A promotional price halves the headline cost advantage, and benchmark configurations (the exact setup under which models are tested) rarely match any single customer's real workload. DeepSWE, Terminal-Bench and SWE Atlas also ran on Microsoft's internal MSBench platform through GitHub Copilot CLI without subagents — meaning the harness itself was Microsoft's. In plain words: the numbers are plausible, but they were measured under conditions Microsoft chose, at a discount Microsoft controls.

How does the Flash variant compare?

For MAI Code 2 Flash, Microsoft reports a 61.0% average pass rate and $0.052 average cost per task across three benchmarks: DeepSWE 1.1, Terminal-Bench 2.1 and SWE Atlas. GPT-6 Luna averaged 55.1% at $0.080 per task. Claude Haiku 5.5 averaged 53.9% at $0.319.

Here the claim is stronger than near-parity. Flash reports both a higher average pass rate and a lower cost per task than either comparator in its class. At $0.052 per task, Flash costs roughly a sixth of Claude Haiku 5.5's reported price while scoring seven points higher on average.

The same caveats apply: promotional pricing, Microsoft's own benchmark harness and a narrow workload definition. Whether the advantage survives standard list pricing is the question the announcement does not answer.

What about user interface work?

Separate from the automated benchmarks, Microsoft ran blind frontend evaluations covering user interfaces, games, data visualizations, 3D scenes and SVG graphics. The company reports that 258 annotators preferred MAI Code 2 over GPT-6 Sol 67% of the time on aesthetics and functionality.

A 67% preference rate across 258 human raters is a substantial margin. Blind evaluation (annotators not knowing which model produced which output) is a reasonable methodology, though the annotator pool and task selection were again Microsoft's. In plain words: human reviewers liked the output more, under a test Microsoft designed.

Where can customers actually use the models?

Both models are available now in GitHub Copilot. MAI Code 2 Flash is also in limited-access preview in Microsoft Foundry, with releases planned for OpenRouter, Vercel, Hermes, OpenCode and Kilo.

The distribution footprint matters as much as the benchmark table. GitHub Copilot already sits inside the workflows of millions of developers, and the planned third-party routes — OpenRouter and Vercel in particular — extend the models beyond Microsoft's own surfaces. A model that ships pre-installed does not need to win every benchmark to win share.

How far beyond coding does this reach?

Microsoft has also optimized Code 2 Flash into a derivative called MAI Work Flash for use across Microsoft 365 and Fabric IQ. This extends the release beyond coding into everyday enterprise work.

On internal Excel customer-scenario benchmarks, Microsoft reports comparable accuracy and token efficiency with GPT-6 Luna, and similar quality and token efficiency on frequent PowerPoint tasks. "Comparable" is the operative word: the derivative matches, rather than beats, its comparator on these office tasks.

In Teams evaluations, the claims are stronger. Microsoft reports higher scores than GPT-6 Luna for decision tracking and knowledge gaps, adapting custom instructions, and tool calling, with up to 28% fewer tokens on instruction following and tool calling.

What does MAI Work Flash do for Power BI?

The most lopsided numbers in the announcement concern analytics. For Power BI semantic models, Microsoft reports a 0.77 overall score for MAI Work Flash in offline DAX query-generation evaluations, compared with 0.54 for GPT-6 Astra, 0.45 for GPT-5.6 Sol and 0.30 for Opus 5.5.

DAX query generation is the capability behind Power BI's Chat with your Data feature, which lets analysts query business data in natural language. A 0.77-versus-0.54 gap against the nearest comparator is not a rounding error. If it holds in production, it would give Microsoft a concrete edge in a feature enterprises already pay for.

These are offline evaluations on Microsoft's own tasks, so the usual skepticism applies. In plain words: the office-suite numbers are the least independently verifiable claims in the release, and the ones most dependent on internal test design.

Why the pricing framing matters

Microsoft's model release is therefore also a distribution release. Copilot gets two coding options, while a Flash derivative moves into spreadsheets, presentations, chat and analytics.

The reported numbers make cost per task the main competitive lever. For enterprises running millions of coding and office tasks per month, a $0.052 task cost versus $0.319 is the difference between a line item and a rounding error. OpenAI and Anthropic can match the strategy only by cutting prices or shipping cheaper tiers of their own.

The unresolved questions are the ones Microsoft flags itself. The 50% promotional price has an unspecified end date. The benchmarks ran on Microsoft's internal MSBench harness. The models are explicitly not tuned for long-horizon agentic work, which is where much of the industry's attention and spending now sits. Until independent evaluations appear, the price-performance claim rests on the company's own arithmetic — persuasive on its face, and priced to move.

Source: unite.ai

Share this article:

More from James Calloway

James Calloway

Show full bio

News editor covering people, launches and day-to-day practice in artificial intelligence and machine learning at AI In Context.

722 articles

Related articles

  1. Microsoft's MAI Code 2 hits 96% of Claude Opus 5.5 at 5% of the cost
  2. Microsoft’s Ignite 2026 adds Copilot Workspaces and agent controls
  3. Anthropic and OpenAI Ship New Models With the Same Pitch: More for Less
  4. Meta and Microsoft sharply cut Claude usage as Anthropic becomes a rival
  5. Microsoft merges coding and productivity AI into one Copilot app

« Previous articleNext article »