AI Agent - Aug 13, 2026

MAI-Code-1.1-Flash: Vision, Cost and Performance

Quick Answer

MAI-Code-1.1-Flash adds native vision for understanding screenshots, diagrams, and designs. Microsoft reports higher coding quality, 25% faster token streaming, 25% fewer tokens per task, and one quarter of MAI-Code-1-Flash’s cost. GitHub currently lists provider prices of $0.20 input, $0.02 cached input, and $1.20 output per million tokens.

These facts answer different questions. Vision is an input capability. Benchmarks are vendor evidence. Provider token price is not the same as a Copilot premium-request multiplier. Evaluate all three on a representative task set.

Vendor-Reported Evidence

Microsoft reports the following comparisons with the June model:

MeasureReported changeEvaluation caveat
Terminal-Bench 2.1 in Copilot CLI+22%Benchmark and product configuration matter
.NET tasks+15%Language and task mix may differ from yours
Code survival+4%Production telemetry; definition and cohort are provider-owned
Return visits+9%Engagement is not code correctness
Token streaming25% fasterPerceived latency also depends on client, tools, and network
Tokens to complete a task25% fewerCompleted-task quality still matters

Do not add the percentages into a composite score. Reproduce the tasks relevant to the repository.

Native Vision Test

Microsoft describes screenshot-, diagram-, and design-understanding workflows. GitHub describes native vision support. Test it with three controlled inputs:

  1. a screenshot with a known layout and accessibility defects;
  2. a diagram with explicit relationships and deliberately ambiguous labels;
  3. a design paired with repository conventions and acceptance criteria.

Score extraction accuracy, missing details, unsupported assumptions, code correctness, responsive behavior, accessibility, and reviewer edits. Remove secrets and personal data from images. Image understanding does not authorize use of the depicted design or assets.

Two Cost Systems

Provider list price

GitHub’s current model-pricing table lists $0.20 input, $0.02 cached input, and $1.20 output per million tokens for MAI-Code-1.1-Flash. GitHub described this as 73% below MAI-Code-1-Flash; Microsoft summarized it as one quarter of the old price. Both are provider comparisons and should stay dated.

Copilot premium requests

GitHub’s launch states a 0.25× premium request multiplier for annual subscribers. That is not a universal token-price discount. Confirm the plan, annual status, request-based or usage-based billing, included allowance, and overage settings.

Completed-Task Evaluation

Measure:

  • tasks accepted without correction;
  • tests and security checks passed;
  • reviewer minutes and changed lines;
  • total latency, including tools;
  • premium requests or token cost;
  • retries and abandoned outputs;
  • image-input benefit versus text-only context.

Use the migration readiness tool to make the comparison repeatable.

Frequently Asked Questions

Does MAI-Code-1.1-Flash support screenshots?

Yes. Microsoft and GitHub describe native vision for screenshots, diagrams, and designs. Exact input support still depends on the Copilot surface and client.

Is it 25% faster for every coding task?

No. Microsoft reports 25% faster token streaming in GitHub Copilot. Total task time can vary with reasoning, tools, network, retries, tests, and review.

What are the current provider token prices?

GitHub currently lists $0.20 input, $0.02 cached input, and $1.20 output per million tokens. Recheck the live table before budgeting.

Is provider token price the same as the Copilot multiplier?

No. Provider list pricing and Copilot premium-request accounting are separate billing layers. The launch’s 0.25x statement is for annual subscribers.

What is the best success metric?

Use accepted, tested work per total cost and reviewer minute. Token speed or benchmark scores alone do not establish completed-task value.

Official Sources

Next Decision