Microsoft Brings MAI Code 2 to GitHub Copilot at 5% of Claude's Cost
Microsoft's new in-house coding models claim near-frontier quality at a fraction of the price, and they're already live inside GitHub Copilot.
- Microsoft launched MAI Code 2 and MAI Code 2 Flash, its second-gen in-house coding models, inside GitHub Copilot
- MAI Code 2 claims 98% of GPT 6.1 Sol performance at 40% cost, 96% of Claude Opus 5.5 at 5% cost
- Flash matches GPT 6 Luna at 65% of its price and Claude Haiku 5.5 at 20%, targeting high-volume routine work
- Available across Copilot Free through Enterprise tiers via GitHub Copilot, CLI, cloud agent, and mobile
- Explicitly not tuned for fundamental research or very long-horizon agentic workloads, Microsoft warns
- Flash also powers MAI Work Flash, now driving AI features inside Word, Teams, Excel and Fabric IQ
Microsoft has added two in-house coding models to GitHub Copilot, pairing near-frontier benchmark claims with substantially lower reported costs. Microsoft’s announcement positions MAI Code 2 for complex software engineering and MAI Code 2 Flash for fast, high-volume tasks.
GitHub Copilot already offers models from several providers. Adding Microsoft’s second-generation MAI-Code models could reduce the company’s reliance on external systems for common coding workloads. The models also expand beyond code completion into agentic work, where a system plans and executes multiple steps across files, tools, and repository context.
Two models, two cost profiles
| Model | Primary workloads | Copilot plans |
|---|---|---|
| MAI Code 2 | Debugging, multi-file implementation, repository-wide reasoning, production changes, and interface development | Pro, Pro+, Max, Business, and Enterprise |
| MAI Code 2 Flash | Code explanations, targeted edits, routine fixes, rapid iterations, and high-volume agent loops | Free, Student, Pro, Pro+, Max, Business, and Enterprise |
Both models are available through the Copilot CLI, coding agent, Copilot app, GitHub.com, and GitHub Mobile. Flash also has a limited-access preview in Microsoft Foundry. Microsoft says support through OpenRouter, Vercel, Hermes, OpenCode, and Kilo will follow.
Business and Enterprise administrators can control access through Copilot’s model policy settings. Those settings govern model availability, while existing repository permissions, retention rules, and compliance controls continue to apply.
Four benchmarks carry the price claim
Microsoft averaged results from DeepSWE 1.1, Terminal Bench 2.1, SWE-Atlas, and FrontierCode v1.1. These evaluations cover repository and terminal-based tasks that more closely resemble software engineering work than isolated code-completion tests.
| Microsoft model | Comparison model | Reported performance | Reported cost |
|---|---|---|---|
| MAI Code 2 | GPT-6.1 Sol | 98% | 40% |
| MAI Code 2 | Claude Opus 5.5 | 96% | 5% |
| MAI Code 2 Flash | GPT-6 Luna | Comparable | 65% |
| MAI Code 2 Flash | Claude Haiku 5.5 | Comparable | 20% |
An evaluation conducted at AT&T reportedly found that Flash outperformed certain GPT-6.1 variants at roughly one-fiftieth of the cost. That result covered specific variants and should not be generalized to the full GPT-6.1 family.
The comparisons come from Microsoft and have not received independent verification. Their cost calculations also use a 50% promotional price, so the ratios will change after the discount expires unless Microsoft adjusts its standard rates.
Code 2 leans into production work
Microsoft identifies debugging, multi-file changes, repository-wide reasoning, and interface development as MAI Code 2’s strongest areas. Its focus on producing reviewable changes makes it the more suitable option for feature implementation and refactoring that spans several parts of a codebase.
In blind comparisons with GPT-6 Sol, evaluators preferred MAI Code 2’s output in 67% of tests covering interfaces, games, data visualizations, 3D scenes, and SVG graphics. Microsoft says 258 annotators participated. The result measures visual preference, while correctness, accessibility, maintainability, and runtime performance require separate evaluation.
Microsoft positions Flash as the smaller and faster option for repetitive work. Its likely fit includes CI assistants, code-review automation, explanation tools, and agent loops that issue many narrow requests, where latency and per-call cost accumulate quickly.
The tuning has clear limits
Microsoft says the models have limited tuning for fundamental scientific research and very long-horizon agentic tasks. Long-horizon work requires a model to maintain context, revise plans, and recover from errors across extended sequences of actions. Teams building autonomous workflows should test those behaviors directly and retain human review for consequential changes.
Flash already reaches beyond Copilot
Microsoft has not disclosed model size, architecture, context-window limits, training data, or inference requirements. It has revealed that MAI Code 2 Flash serves as the base for MAI Work Flash, a derivative optimized for Microsoft 365 and Fabric IQ.
- Word: Microsoft reports that its model performs similar to GPT 6.1 Sol on 5 key internal benchmarks while using approximately 42% fewer input tokens per call in production.
- Teams: The company reports better results than GPT-6 Luna for tracking decisions and following custom instructions, with up to 28% fewer tokens.
- Excel and PowerPoint: Internal customer-scenario benchmarks show accuracy and token efficiency comparable to GPT-6 Luna.
- Fabric IQ: Microsoft reports stronger DAX query generation than GPT-5.6 Sol, GPT-6 Astra, and Claude Opus 5.5. DAX, or Data Analysis Expressions, is the query and formula language used by Power BI.
These internal deployments show how Microsoft can adapt one base model to several product domains. The results remain vendor-reported, and the company has not published enough detail for outside reproduction.
Microsoft trims its model dependence
Copilot has historically depended on models supplied by OpenAI and other AI companies. Competitive in-house models give Microsoft more control over inference costs, product tuning, deployment schedules, and capacity. They also let GitHub route suitable workloads to Microsoft models while preserving larger external systems for tasks that need them.
The release accompanies a broader expansion of the MAI portfolio. MAI Transcribe 2 has launched in Dragon Copilot for medical workflows, while MAI Cyber 2 Flash is planned for Microsoft Security products. Microsoft has also introduced image and real-time voice models.
Microsoft is separately licensing model weights, allowing eligible enterprises to customize variants and deploy them in Azure environments under the applicable terms. That option gives teams more control than a managed API alone, although licensing, infrastructure, security, and maintenance requirements still need review.
How to evaluate the switch
- Build a representative task set. Include bug fixes, refactors, test generation, interface work, and repository-level changes drawn from your normal workload.
- Measure completed work. Track test pass rates, reviewer corrections, latency, token use, failed tool calls, and total cost per accepted change.
- Use MAI Code 2 for broader changes. Its stated strengths align with debugging, multi-file implementation, and visual interface work.
- Use Flash for narrow, frequent calls. Explanations, targeted edits, CI automation, and agent loops benefit most from its lower reported cost.
- Keep alternatives available for extended planning. Microsoft identifies long-horizon work as a weaker area, so demanding autonomous tasks need direct comparison with larger models.
- Check organizational policy. Enterprise users should confirm model access, data controls, repository permissions, and post-promotion pricing before rollout.
For existing Copilot customers, the practical next step is a controlled comparison on real repositories. Microsoft’s figures suggest a strong cost curve for routine enterprise development, but independent benchmarks and production measurements will determine whether the savings survive code review, retries, and the end of promotional pricing.