Released August 11, 2026 as a mid-cycle refresh, MAI-Code-1.1-Flash targets everyday GitHub Copilot workflows rather than frontier-scale reasoning, trading some raw benchmark ceiling for Microsoft's tightest first-party integration. Terminal-Bench performance climbed 22% over its June predecessor, making it the pick for teams already living inside Copilot's CLI and IDE surfaces.
MAI-Code-1.1-Flash is Microsoft's small-tier coding model for GitHub Copilot, running a sparse Mixture-of-Experts design that pairs 137B total parameters with 5B active per token. It scores 72.6% on SWE-bench Verified and adds native vision input, a capability its text-only predecessor lacked. A 256,000-token context window handles large repositories in a single pass.
Provider: Microsoft · Family: MAI-Code
Context window: 256,000 tokens · Max output: 128,000
Input modalities: text, image, tool-calls · Output: text, code, tool-calls
About MAI-Code-1.1-Flash
MAI-Code-1.1-Flash is Microsoft's in-house small-tier coding model, released August 11, 2026 as the successor to MAI-Code-1-Flash (released June 2, 2026). It runs on a sparse Mixture-of-Experts architecture with 137B total parameters and 5B active parameters per token. It sits inside Microsoft's MAI model family alongside MAI-Thinking-1, which handles the deliberate reasoning work this model does not attempt, and it exists specifically to be GitHub Copilot's low-cost, high-volume coding workhorse rather than a general frontier model. On benchmarks, MAI-Code-1.1-Flash scores 72.6% on SWE-bench Verified, averaging 8.6K tokens per solved task, and 62.9% on Terminal-Bench 2.1 inside GitHub Copilot CLI, averaging 17.0K tokens. That result is a 22% improvement over the predecessor's score on the prior terminal benchmark, and Microsoft also reports a 15% gain on .NET-specific tasks. The predecessor already beat Claude Haiku 4.5 on SWE-bench Verified (71.6% vs 66.6%) and SWE-bench Pro (51.2% vs 35.2%), and 1.1 extends that lead while using roughly 25% fewer tokens per solve than the June release did. Context window sits at 256,000 tokens with a 128,000-token maximum output ceiling, unchanged from the predecessor. Microsoft reports 25% faster token streaming inside GitHub Copilot's CLI, which combined with the lower token count per solve shortens the wall-clock time of longer agentic sessions. Inputs are text and, new in this release, native images: MAI-Code-1.1-Flash can read screenshots, diagrams, and UI mockups directly, a capability the text-only June release lacked. Output stays text and code, with native function calling for tool use. There is no toggleable extended-reasoning mode; Microsoft routes deliberate multi-step planning to the separate MAI-Thinking-1 model, keeping this one tuned for fast, single-pass code generation, repository question answering, and refactoring. Pricing dropped sharply from the June release: Microsoft cut the rate across every tier of the per-token structure, framing the update as delivering more capability for meaningfully less spend. The model bills through GitHub Copilot's premium-request system and through Microsoft Foundry directly, with annual Copilot subscribers charged at a fraction of a standard premium request per call. Exact current rates and worked cost examples live in the pricing FAQ below. MAI-Code-1.1-Flash is live across VS Code, Visual Studio, JetBrains IDEs, Eclipse, Xcode, GitHub Mobile, Copilot CLI, the Copilot cloud agent, and Copilot Chat on GitHub, spanning every Copilot plan from Free through Enterprise. Free and Student users get it only through automatic model selection; paid tiers can select it manually. Business and Enterprise administrators have to switch on its Copilot policy themselves, since it ships off by default at the organization level. Outside Copilot it is reachable through Microsoft Foundry on Azure and third-party routers including OpenRouter, Fireworks AI, and Baseten; no AWS Bedrock or Google Vertex listing has surfaced. On safety, Microsoft's published model card describes filtering harmful content out of the pre-training mixture, then applying supervised fine-tuning and reinforcement learning to shape model behavior toward safe, helpful output. Weights are closed: there is no Hugging Face release or open license, and Microsoft has not disclosed a specific training data cutoff date for this 1.1 refresh. The model fits teams already running on GitHub Copilot who want an inexpensive, fast agent for CLI-heavy or .NET-heavy work, and for high-volume tasks like bulk refactors or repo Q&A where a frontier-priced model is overkill. Teams that need deliberate, multi-step reasoning are better served by MAI-Thinking-1 or a frontier coding model, and DeepSeek-V4-Flash-0731 posts a stronger terminal-benchmark score at a lower blended price for teams not tied to the Copilot ecosystem. MAI-Code-1-Flash, the June predecessor, is being deprecated across all GitHub Copilot surfaces on September 10, 2026, and Microsoft has described 1.1 as a drop-in replacement at the prompt level rather than a rewrite.
Pricing
Through GitHub Copilot and Microsoft Foundry, input runs $0.20 per 1M tokens, cached input $0.02 per 1M, and output $1.20 per 1M, a 73.3% cut across the board versus MAI-Code-1-Flash's $0.75/$0.075/$4.50 rates. Annual Copilot subscribers pay a 0.25x premium request multiplier.
Key Features
- Native Vision Input: New in 1.1: the model reads UI mockups, diagrams, and error screenshots on its own, something the June text-only release couldn't do.
- 256K Context Window: Holds an entire mid-size repository or a long multi-file diff in context, with a 128K token output ceiling for long generated diffs or docs.
- Agentic Terminal Coding: Tuned against Copilot CLI's own reinforcement-learning environments, where it posts its strongest score on the terminal-coding benchmark.
- Prompt Caching: Cached input tokens cost a small fraction of the standard input rate, a real discount for repeat-context agent loops. Exact rates are in the pricing FAQ.
- Native Function Calling: Built-in tool use and structured output support that plugs directly into Copilot's agent loop without a separate function-calling adapter.
Pros
- Cheapest first-party model Microsoft has put in Copilot to date, undercutting even its own June release by a wide margin.
- Available across nine Copilot surfaces on day one, from VS Code and Visual Studio to Copilot CLI and the cloud agent.
- Vision input lets it debug from a screenshot instead of a text description, closing a gap the June release had.
Cons
- DeepSeek-V4-Flash-0731 posts a stronger terminal-benchmark result at a lower blended token price, so it is not the strongest option on raw benchmarks alone.
- Closed weights with no Hugging Face release; teams that need to self-host or fine-tune the base model have to look elsewhere.
- Requires manual admin setup before Business and Enterprise developers can even select it.
Benchmarks
- swe bench verified: 72.6
- terminal bench 2 1: 62.9
Frequently Asked Questions
How much does MAI-Code-1.1-Flash cost per 1M tokens?
MAI-Code-1.1-Flash costs $0.20 per 1M input tokens, $0.02 per 1M cached input tokens, and $1.20 per 1M output tokens through GitHub Copilot and Microsoft Foundry. That's a 73.3% cut across all three rates versus MAI-Code-1-Flash's $0.75/$0.075/$4.50. Annual Copilot subscribers also pay a 0.25x premium request multiplier on top of their base plan.
How does MAI-Code-1.1-Flash compare on benchmarks vs DeepSeek-V4-Flash-0731?
DeepSeek-V4-Flash-0731 posts a stronger 82.7% pass rate on the terminal-coding benchmark at a lower blended token price than Microsoft's model manages on that same eval, per comparisons published after both launched. Microsoft has not published a head-to-head SWE-bench Verified score for DeepSeek's model, so that axis remains a Microsoft-only claim; DeepSeek is the better raw-benchmark pick for teams not tied to GitHub Copilot.
Is MAI-Code-1.1-Flash open source or proprietary?
MAI-Code-1.1-Flash is fully proprietary with closed weights; Microsoft has not published it on Hugging Face or under any open license. Access runs through GitHub Copilot, from the Free plan through Enterprise, or through Microsoft Foundry, plus third-party routers including OpenRouter, Fireworks AI, and Baseten.
Does MAI-Code-1.1-Flash train on user data?
Microsoft has not published a specific data-retention or training-on-inputs policy for this release. Its model card outlines pre-training content filters plus a fine-tuning and reinforcement-learning stage aimed at safe, helpful behavior, but does not state whether Copilot request data feeds back into future training runs.
Who is MAI-Code-1.1-Flash best for and who should avoid it?
It's the right call for engineering teams standing on GitHub Copilot already, chasing a cheap, quick model for CLI- or .NET-heavy agent work, especially now that native vision input is built in. Teams needing deliberate multi-step reasoning should look at MAI-Thinking-1 instead, and teams chasing the strongest raw coding benchmark at a similar price should compare DeepSeek-V4-Flash-0731 before committing.
Top Alternatives
- DeepSeek-V4-Flash-0731: Pick DeepSeek-V4-Flash-0731 if terminal-benchmark performance matters more than native Copilot integration; its 82.7% beats this model's 62.9% at a lower per-token rate.
- Kimi K2.7-Code: Pick Kimi K2.7-Code if you need open weights to self-host; it costs more per token than MAI-Code-1.1-Flash but isn't locked to Microsoft's Copilot ecosystem.
- GPT-5.6 Luna: Pick GPT-5.6 Luna if you want OpenAI's budget-tier model outside the Copilot toolchain; pick MAI-Code-1.1-Flash if your team already lives inside GitHub Copilot.