# Microsoft upgrades its Copilot coding model 10 weeks after launch > Source: > Published: 2026-08-11 18:40:01+00:00 Microsoft released MAI-Code-1.1-Flash on August 11th and put the coding model into production in GitHub Copilot, replacing the first version of its in-house coding model just 10 weeks after launch. In its [announcement](https://microsoft.ai/news/mai-code-1-1-flash-br-better-faster-at-a-quarter-of-the-cost/), Microsoft said the update streams tokens 25% faster, uses 25% fewer tokens to finish a task and costs one quarter as much as MAI-Code-1-Flash, which debuted on June 2nd. Microsoft also added image input, allowing the model to turn screenshots, diagrams and interface designs into code. The unusually short release cycle shows how Microsoft plans to compete in coding models: use Copilot as both the distribution channel and the training environment, then rapidly replace models as production data reveals where developers struggle. The company trained the new version around command-line and .NET work after feedback identified both as weak points. ### A large model with a small active footprint The [MAI-Code-1.1-Flash model card](https://microsoft.ai/pdf/MAI-Code-1.1-Flash-Model-Card.PDF) describes a sparse mixture-of-experts model with 138 billion total parameters and 5 billion active parameters. It accepts text and images, has a 256,000-token context window and was trained between March and August 2026. That design lets Microsoft keep the model's active compute requirements closer to a small model while retaining a much larger pool of parameters. MAI-Code-1.1-Flash starts from a compressed, 5 billion-active-parameter checkpoint of MAI-Thinking-1 and depends on the original coding model. Microsoft used roughly 2 million synthetic agent tasks during an intermediate training stage, followed by reinforcement learning across more than 150,000 environments. The company evaluated the resulting checkpoints inside the same GitHub Copilot harness used for production traffic, including repository navigation, tool calls and end-to-end task verification. The model's clearest benchmark gain came on Terminal-Bench 2.1, which measures an agent's ability to complete work through a terminal. MAI-Code-1.1-Flash recorded a 62.9% pass rate, up from 51.7% for its predecessor. That is the 22% relative improvement cited by Microsoft. The gain on SWE-Bench Verified was narrower. The new model scored 72.6%, compared with 71.6% for the June model. It used an average of 8,600 solution tokens per task, down from 10,800. Microsoft's results also put it above GPT-5.4 mini on both tests, although the figures were generated using Microsoft's Copilot production harness rather than a neutral third-party evaluation. Microsoft separately reported a 15% improvement on .NET tasks. Its production measures showed 4% more generated code surviving through a commit and a 9% increase in developer return visits. Those figures are internal, self-reported metrics, and the announcement does not specify their baseline or measurement window. ### Copilot data feeds the next model The model update gives a more concrete view of the training loop Microsoft outlined when it [introduced MAI-Code-1-Flash in June](https://microsoft.ai/news/introducingmai-code-1-flash/). Microsoft can train inside the Copilot harness, route the model to developers, observe aggregate product outcomes and focus the next training run on tasks that depress acceptance or retention. The [data summary](https://microsoft.ai/pdf/MAI-Code-1.1-Flash-Data-Card.PDF) says Microsoft used conversation prompts and associated context from GitHub Copilot Free, Pro and Pro+ subscribers who had not opted out of model training. The filtered data was used for reinforcement-learning rollouts and to train an internal reward model. Microsoft said it removed identifying information, images, tool-use traces and safety-sensitive content, and did not use production responses as supervised training targets. Training data included more than 10 trillion text tokens and fewer than 1 million images, according to the summary. The corpus combined public repositories, web data, books, academic material, commercially acquired data and synthetic software-engineering tasks. Its latest collected datasets dated to July 2026, a month before release. This production feedback loop is central to the economics of the release. Microsoft said in July that millions of developers were already using the first MAI-Code model and that it produced higher code acceptance and return rates than GPT-5.4 mini and Claude Haiku 4.5. Moving more Copilot traffic onto a cheaper Microsoft-built model can reduce the cost of serving high-volume agent sessions while giving Microsoft tighter control over latency, training and product-specific behavior. Pricing remains uneven across Microsoft's release materials. The model card marks pricing as "to be finalized," while GitHub's [live model billing table](https://docs.github.com/en/copilot/reference/copilot-billing/models-and-pricing) currently lists the broader MAI-Code-1-Flash entry at $0.75 per million input tokens, $0.075 for cached input and $4.50 per million output tokens. Microsoft's announcement does not attach specific dollar rates to the claimed 75% reduction. MAI-Code-1.1-Flash is available through [GitHub Copilot](https://github.com/features/copilot). Microsoft has also opened an [official repository](https://github.com/microsoft/MAI-Code) for release information and developer feedback.