MAI-Code-1.1-Flash: Copilot and Local Coding Costs

Select MAI-Code-1.1-Flash in the GitHub Copilot model picker when it is enabled for your plan and organization. GitHub’s August 11 release marks the cloud model as available, and its current pricing table lists it as GA. An organization’s model policy can affect access.

The cloud model accepts text and images and returns text, with a published 256K-token context. Use a cropped screenshot with the relevant repository files when implementing a UI change. Ask for a bounded change and a test plan before giving an agent a long sequence of editing tasks.

The model card applies the relevant GitHub or Microsoft product and service terms. Download availability for the local build is separate from an Apache or MIT weight license. Use Microsoft’s documented local distribution and its terms rather than treating the name as evidence of unrestricted redistribution.

GitHub’s usage-based model table, checked October 9, 2026, quotes $0.20 per million input tokens, $0.02 per million cached-input tokens and $1.20 per million output tokens. These are Copilot model rates. The cloud route is hosted on Azure in GitHub’s tenant; they are not a generic Azure endpoint quote.

An assumed request with 50,000 uncached input tokens and 5,000 output tokens costs $0.010 + $0.006 = $0.016 at those rates. If all 50,000 input tokens qualify as cached input, that line becomes $0.001 and the total $0.007. Add every agent turn, tool result and retry to the task total; the first prompt alone is not the whole bill.

GitHub’s release notice distinguishes usage-based billing from the 0.25× premium-request multiplier for annual Copilot Pro and Pro+ subscribers on request-based billing. Apply the billing scheme on your account. Usage-based billing converts the token total to AI credits at 1 credit = $0.01; the $0.016 example is 1.6 credits. Keep the plan’s included allowance and additional usage conditions beside that total. Do not multiply a token bill by the old request multiplier.

Microsoft’s October 7 local announcement recommends more than 120GB of RAM for best performance and says the model can be downloaded and run locally. It separately schedules experimental Copilot app, CLI and VS Code integration by the end of October. Treat local distribution and that integrated experience as separate availability milestones.

The local engineering post describes 137B total and 6.8B active parameters, a 53GB quantized footprint and 75.5GB peak memory at 256K context on Surface Laptop Ultra. Its test used approximately 3.3-bit mixed precision, DFlash2 speculative decoding and Windows ARM64 llama.cpp CUDA on October 5. These are that configuration’s measurements, not minimum requirements for every PC.

The earlier cloud model card lists 138B total and 5B active parameters. Keep those cloud figures separate from the local post’s specifications when choosing an artifact. Active parameters reduce per-token computation; the device still needs storage and memory for the full selected model, its cache, runtime and drafter.

Local model calls carry no per-inference charge in Microsoft’s announcement. Budget electricity, device ownership, setup time and any cloud work selected by the agent. Local model selection, cloud routing and tool execution are different boundaries; choosing a local model does not make every tool or session offline.

At an illustrative 250W wall draw, two hours per day and $0.15/kWh, electricity costs $2.25 per month. If a $3,600 machine is allocated over 36 months, hardware adds $100 per month before resale value or other workloads.

At the $0.016 cloud request example above, $100 buys 6,250 such requests before plan credits, tools and retries. This comparison only helps once you know the actual token mix and accepted-task rate. An already-owned workstation and a newly purchased workstation have different economics.

Use the same repository snapshot, issue, tool permissions and passing tests across models. Record the accepted patch, correction turns, review minutes and complete cloud usage or local wall time. A screenshot-to-code task also needs a visual check of the requested layout after the tests pass.

Microsoft’s local post reports SWE-Bench Verified and Terminal-Bench 2.1 on 500 and 89 cases, respectively, for its stated cloud and quantized configurations. The local scores are 70.80% and 66.29%. Keep those publisher measurements with their hardware and harness context; they do not rank a different Copilot setup against Sol or Qwen.

Begin with repository Q&A, a constrained refactor or a UI change that has clear acceptance criteria. Increase autonomy after the model has shown it can select tools, recover from errors and produce changes you can review in that environment. Local inference does not constrain shell access by itself: use the harness’s sandbox and limit files, network access and credentials to the task.