Back Home

最新模型

Microsoft Releases MAI Code 2 Series, with Coding Model Cost Comparison Based on Promotional Pricing

Microsoft released MAI Code 2 and Code 2 Flash on November 17, making them available in GitHub Copilot. The new models target cost-effective enterprise coding, but the official comparison uses promotional pricing and specific agent settings, so results should be validated against real workloads.

Mike Mozart · CC BY 2.0 · Image source
zh-Hant

On November 17, Microsoft released MAI Code 2 and MAI Code 2 Flash, bringing its in-house coding models to GitHub Copilot. The larger Code 2 is aimed at deep debugging, multi-file implementation, cross-repository reasoning, and interface development. Flash targets code explanation, localized edits, and frequent patches. The launch post also says Flash is available in limited preview in Microsoft Foundry, while integrations with other platforms remain planned for the future. Official announcement

The model card discloses that both models use a decoder-only Transformer with sparse Mixture-of-Experts layers and an integrated vision encoder, accepting text and image inputs. The stated total context length is 1.05 million tokens, including an input limit of 922,000 and an output limit of 128,000 tokens. This allows screenshots, interface requirements, and large amounts of code to enter the same workflow; however, context capacity alone does not prove that a model can reliably complete engineering tasks of equivalent length. Model card

The evaluation conditions deserve close scrutiny. Microsoft says Code 2 achieves an average of 98% of GPT 6.1 Sol’s performance across four coding benchmarks, at 40% of the cost. DeepSWE 1.1, TerminalBench 2.1, and SWE Atlas were run on Microsoft’s internal MSBench platform with Copilot CLI, with subagents disabled. For FrontierCode 1.1, Microsoft cites scores and costs from Artificial Analysis. All models were evaluated at the highest reasoning effort, and Code 2’s cost was calculated using a 50% promotional discount. These figures therefore should not be treated as standard API pricing or as the savings a team will realize after deployment. Evaluation methodology and pricing notes

For engineering teams, the split between the larger and smaller models offers an option to assign models by task: teams can start by testing Flash for routine changes, then evaluate Code 2 for issues spanning multiple files. Microsoft also confirmed in its Ignite announcement that both models are now in the Copilot model picker and Fusion. Next, teams should track costs after the promotion ends, patch success rates across different tool frameworks, and the effect of long contexts on latency. Microsoft also states that the series is not yet optimized for fundamental scientific research or very long-running agent tasks. Adoption should therefore be judged using test pass rates on a team’s own repositories and the results of human review. Official Ignite announcement

Sources

  1. MAI Code 2 and Code 2 Flash: Everyday coding at scalable prices
  2. MAI Code 2 and MAI Code 2 Flash Model Card
  3. Ignite 2026: Building the frontier ecosystem