Skip to the cards
Filed
Source
Michael Davis
Loans
966

Microsoft MAI Code 2 Models: Benchmark Scores, Pricing and Enterprise Checks

MAI Code 2 trails top rivals on average benchmarks, while Flash leads its comparison group. Promotional token prices run through May 17, 2027.

Microsoft MAI Code 2 Models: Benchmark Scores, Pricing and Enterprise Checks

Microsoft’s MAI Code 2 models give development teams another option inside GitHub Copilot, with different models aimed at complex repository work and frequent, smaller coding tasks. The benchmark results show a mixed picture: MAI Code 2 sits close to leading rivals overall, while MAI Code 2 Flash leads its comparison group on average. The pricing advantage needs careful reading, because Microsoft’s cost comparisons use promotional rates for its own models.

Both models launched at Ignite 2026. Microsoft says they focus on routine enterprise development needs rather than fundamental science research or long-running agent tasks. That scope matters when teams evaluate the models: the results are most relevant to the coding work and benchmark conditions used in the comparisons, not every kind of software engineering.

Two models for different coding workloads

MAI Code 2 is intended for work that may involve several connected steps: deep debugging, changes across multiple files, reasoning across a repository, and user interface development. MAI Code 2 Flash is aimed at higher-volume requests, including explaining code, making targeted edits, handling routine fixes, and supporting quick iterations.

That split gives teams a basis for testing whether a faster, lower-cost model can handle common requests while reserving the larger model for tasks that require broader context. Microsoft’s description does not position either model as tuned for fundamental science or long-horizon agentic work. Teams with those workloads should treat the published coding comparisons as limited evidence for their needs.

Microsoft says it developed the models around the coding work its customers do each day. The launch also fits the company’s effort to build more of its own AI capabilities. Pareekh Jain, CEO at EIIRTrend & Pareekh Consulting, said Microsoft can connect its models with GitHub, Visual Studio, and Azure, platforms already used by developers. He described reducing AI costs and strengthening Microsoft’s developer ecosystem as immediate objectives.

MAI Code 2 is close on average, with uneven results by test

Microsoft compared MAI Code 2 with OpenAI’s GPT-6.1 Sol and Anthropic’s Claude Opus 5.5 across four benchmarks used in GitHub Copilot: DeepSWE 1.1, Terminal-Bench 2.1, SWE-Atlas, and FrontierCode v1.1. The average pass rate was 63.2% for MAI Code 2, 64.1% for GPT-6.1 Sol, and 65.3% for Claude Opus 5.5.

The individual results varied. MAI Code 2 scored 73.2% on DeepSWE 1.1 and 84.6% on Terminal-Bench 2.1; GPT-6.1 Sol scored 74.6% and 85%, while Claude Opus 5.5 scored 69% and 78.7%. On SWE-Atlas, MAI Code 2 reached 43.1%, compared with 47.4% for GPT-6.1 Sol and 57.5% for Claude Opus 5.5. On FrontierCode v1.1, MAI Code 2 scored 52%, ahead of GPT-6.1 Sol at 49.4% but below Claude Opus 5.5 at 56%.

Those figures put MAI Code 2 in range on some tests but behind on others. For teams, the useful takeaway is to examine the tasks represented by each benchmark and compare them with their own codebase and development workflow. A single average can hide differences that matter to particular projects.

Flash leads its comparison group on average

Microsoft compared MAI Code 2 Flash with GPT-6 Luna and Claude Haiku 5.5 on DeepSWE 1.1, Terminal-Bench 2.1, and SWE-Atlas at maximum reasoning. Flash recorded a 61% average pass rate, compared with 55.1% for GPT-6 Luna and 53.9% for Claude Haiku 5.5.

Flash scored 59.9% on DeepSWE 1.1, 78.3% on Terminal-Bench 2.1, and 44.8% on SWE-Atlas. GPT-6 Luna returned 59.6%, 69.3%, and 36.4%; Claude Haiku 5.5 returned 46%, 73.4%, and 42.1%. Flash therefore led on the average and on two of the three listed tests, while Claude Haiku 5.5 scored higher on Terminal-Bench 2.1.

Promotional rates change the cost comparison

Microsoft reported an average benchmark task cost of $0.309 for MAI Code 2, against $0.761 for GPT-6.1 Sol and $6.306 for Claude Opus 5.5. For Flash, its reported average was $0.052 per task, compared with $0.080 for GPT-6 Luna and $0.319 for Claude Haiku 5.5. Microsoft’s MAI Code 2 figures use 50% promotional pricing, and the benchmark task cost is distinct from a model’s published token rates.

According to GitHub’s pricing page, promotional MAI Code 2 rates are $0.50 per million input tokens, $0.025 per million cached input tokens, and $2.50 per million output tokens through May 17, 2027. After that date, the listed rates rise to $1.00, $0.05, and $5.00, respectively.

MAI Code 2 Flash’s promotional rates are $0.05 per million input tokens, $0.005 per million cached input tokens, and $0.25 per million output tokens. The listed rates after the promotion are $0.10, $0.01, and $0.50, respectively. Sanchit Vir Gogia, chief analyst at Greyhound Research, said the promotional rates narrow the apparent cost gap once they expire, and cautioned that benchmark-run spend is not the same measure as published per-token pricing.

Measure delivery costs, not just model usage

Token charges are only one part of the cost of AI-assisted development. Jain pointed to integration, testing, security review, human review, and correction of generated errors as other expenses. His example: if model usage represents 10% of development costs, a 60% reduction in model costs would translate to 6% savings across total costs, assuming the other expenses stay the same.

Gogia recommends testing models on a team’s own code and tools, using the effort setting intended for production. He also recommends tracking cost per merged code change, change failure rates, and security findings against a baseline from teams not using AI coding models. These measures can show whether cheaper generation translates into more reliable delivery or simply shifts work into review and repair.

Gogia cited Google’s DORA research, which has associated wider AI adoption with weaker delivery stability. He also pointed to model cards that rate both MAI models highly on offensive cyber capabilities. For a pilot, teams should therefore assess generated code quality alongside security and reliability when changes enter production systems. The practical decision is whether each model improves the team’s full path from request to working, reviewed code at the rates and settings it will actually use.

Source: Michael Davis · www.infoworld.com

Filed alongside