Microsoft AI’s MAI Code 2 Targets Near-Rival Coding at Lower Cost
Microsoft says MAI Code 2 nearly matches top rivals on coding benchmarks at a fraction of the cost, while MAI Code 2 Flash leads two comparators and a derivative enters Microsoft 365 and Fabric IQ.
By Grace Kim
4 min read

What's News
- Microsoft announced MAI Code 2 and MAI Code 2 Flash on November 17, 2026.
- MAI Code 2 posted a 63.2% average coding pass rate at $0.309 per task across four benchmarks.
- MAI Code 2 Flash posted a 61.0% average pass rate at $0.052 per task across three benchmarks.
- Microsoft used a 50% promotional price for its models’ reported task costs.
- MAI Work Flash scored 0.77 in offline DAX generation, compared with 0.30 for Opus 5.5.
By Grace Kim
Microsoft reports that MAI Code 2 delivered 98% of GPT-6.1 Sol’s performance at 40% of GPT-6.1 Sol’s cost and 96% of Claude Opus 5.5’s performance at 5% of Claude Opus 5.5’s cost across four coding benchmarks. On November 17, 2026, Microsoft AI announced MAI Code 2 and MAI Code 2 Flash for everyday enterprise coding, with cost per task—not raw benchmark superiority—at the center of the release.
What does Microsoft’s split target?
Microsoft designed MAI Code 2 for production changes, deep debugging, multi-file implementation, repository-wide reasoning and user interfaces. The company aimed MAI Code 2 Flash at high-volume jobs: explaining code, making targeted edits, applying routine fixes and iterating quickly. Microsoft said the split separates deeper implementation work from frequent tasks where lower per-task costs matter at scale.
Microsoft also set a clear boundary: it has not tuned either model for fundamental science research or very long-horizon agentic work. Microsoft said that limitation keeps the release focused on the coding work customers do every day.
For enterprise buyers, Microsoft’s two-tier launch turns model choice into a trade-off between implementation depth and unit cost.
How close are the coding benchmarks?
Across DeepSWE 1.1, Terminal-Bench 2.1, SWE Atlas and FrontierCode v1.1, Microsoft reported 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, while Claude Opus 5.5 averaged 65.3% at $6.306. Microsoft’s results put Code 2 close to both comparators on pass rate while showing a much larger cost gap.
Microsoft said all models in the coding comparisons ran at maximum reasoning effort. Its cost figures for the Microsoft models use a 50% promotional price. Microsoft cautioned that workload, model configuration, reasoning effort or pricing assumptions can change the comparisons.
Microsoft said DeepSWE, Terminal-Bench and SWE Atlas ran on its internal MSBench platform through GitHub Copilot CLI without subagents.
On Microsoft’s numbers, cost—not the top pass rate—is MAI Code 2’s clearest advantage.
Does MAI Code 2 Flash lead both comparators?
Across DeepSWE 1.1, Terminal-Bench 2.1 and SWE Atlas, Microsoft reported a 61.0% average pass rate and $0.052 average cost per task for MAI Code 2 Flash. GPT-6 Luna averaged 55.1% at $0.080, and Claude Haiku 5.5 averaged 53.9% at $0.319. Flash posted both a higher average pass rate and a lower reported task cost than either comparator.
Microsoft also reported results from separate blind frontend evaluations covering user interfaces, games, data visualizations, 3D scenes and SVG graphics. In those tests, 258 annotators preferred MAI Code 2 over GPT-6 Sol 67% of the time on aesthetics and functionality.
Flash leads the reported pass-rate and cost comparison; Code 2 took the majority preference in the separate frontend test.
Where will the models reach customers?
Microsoft made both models available in GitHub Copilot. MAI Code 2 Flash also entered a limited-access preview in Microsoft Foundry, with releases planned for OpenRouter, Vercel, Hermes, OpenCode and Kilo.
Microsoft has further optimized Code 2 Flash into MAI Work Flash for use across Microsoft 365 and Fabric IQ. The company reports comparable accuracy and token efficiency with GPT-6 Luna on internal Excel customer-scenario benchmarks. Microsoft also reports similar quality and token efficiency on frequent PowerPoint tasks.
In Teams evaluations, Microsoft said MAI Work Flash scored higher than GPT-6 Luna on decision tracking and knowledge gaps, as well as on adapting custom instructions and tool calling. Microsoft also reported up to 28% fewer tokens on instruction following and tool calling.
For Power BI semantic models, Microsoft reported a 0.77 overall score for MAI Work Flash in offline DAX query-generation evaluations. GPT-6 Astra scored 0.54, GPT-5.6 Sol scored 0.45 and Opus 5.5 scored 0.30, Microsoft said. DAX query generation powers Power BI’s “Chat with your Data” feature.
Microsoft’s rollout now reaches beyond coding, putting a Flash derivative into spreadsheets, presentations, chat and analytics.
What will decide the enterprise case?
Microsoft’s limited-access Foundry preview and planned releases to OpenRouter, Vercel, Hermes, OpenCode and Kilo will broaden the distribution push. For buyers, the company’s caveat sets the next test: workload, model configuration, reasoning effort and pricing assumptions can change the comparisons, including those built on a 50% promotional rate.
Source: unite.ai
More from Grace Kim
Show full bio
Market editor covering people, launches and day-to-day practice in entrepreneurship, startups and venture funding at Business Bearings.
2,301 articles