Is Kimi K3 available in GitHub Copilot?

Yes. GitHub made Kimi K3 generally available in Copilot on August 6, 2026, hosted on Fireworks AI and billed at provider list pricing of $3 per million input tokens, $15 per million output tokens and $0.30 per million cached input tokens under usage-based billing. It rolled out to Copilot Pro, Pro+, Max, Business and Enterprise plans across VS Code, Visual Studio, Copilot CLI and github.com. It is not available on Free or Student plans.

Which Copilot model should I use for agentic coding in 2026?

Use Kimi K3 for long multi-step agentic work where tool-calling reliability matters and $3/$15 per MTok is acceptable. Use MAI-Code-1.1-Flash for high-volume inline completion, quick edits and anything involving screenshots or design images, since it has native image understanding and is the cheapest path. Reserve a frontier model such as Claude Opus 5 at $5/$25 or GPT-5.6 Sol at $5/$30 for architecture decisions and hard debugging where a wrong answer costs more than the tokens.

Quick Answer

Kimi K3 vs MAI-Code-1.1-Flash in Copilot 2026

Q: What is MAI-Code-1.1-Flash in GitHub Copilot?

MAI-Code-1.1-Flash is Microsoft's in-house fast coding model, released August 11, 2026 and rolled out in GitHub Copilot with native image understanding plus improvements to coding quality, instruction following, tool use and performance. Unlike Kimi K3, it reaches Copilot Free and Student users through auto model selection, making it the model most Copilot users will actually run without choosing anything.

Published: August 18, 2026

The Short Answer

GitHub is betting on model choice, not model loyalty. In one week of August 2026 Copilot gained two models with opposite jobs: Kimi K3 (August 6) for capable agentic work at open-weight pricing, and MAI-Code-1.1-Flash (August 11) for fast, cheap, image-aware everyday coding.

	Kimi K3	MAI-Code-1.1-Flash	GPT-5.6 Sol	Claude Opus 5
In Copilot since	Aug 6, 2026	Aug 11, 2026	Available	Available
Price /MTok	$3 / $15	Lowest tier	$5 / $30	$5 / $25
Cached input	$0.30	—	—	$0.50
Plans	Pro, Pro+, Max, Business, Enterprise	Incl. Free + Student (auto-select)	Paid	Paid
Image understanding	❌	✅ native	✅	✅
Open weights	✅ (July 27, 2026)	❌	❌	❌
Hosted by	Fireworks AI	Microsoft	OpenAI	Anthropic
Best for	Agentic multi-step work	Volume + visual input	Hard reasoning	Long-horizon refactors

Prices are provider list rates under Copilot usage-based billing, verified August 18, 2026.

Cost Per Real Task

Headline rates mislead. At a representative 30K input / 5K output coding task:

Model	Cost per task	Relative
MAI-Code-1.1-Flash	lowest tier	baseline
Kimi K3	$0.165	—
Claude Sonnet 5	$0.110	0.67× K3
Claude Opus 5	$0.275	1.7× K3
GPT-5.6 Sol	$0.300	1.8× K3

Note what that reveals: Kimi K3 is not the cheapest capable option in Copilot. Claude Sonnet 5 at $2/$10 undercuts it by a third on the same task shape. K3’s argument is agentic reliability and the fact that its weights are public, not raw price.

Where Each One Wins

Kimi K3 — agentic work you don’t want to babysit

K3’s open weights landed July 27, 2026, and it has since become a default reference point for multi-step agent workloads: consistent tool-calling, stable behaviour across long sessions, and pricing that doesn’t punish a 40-step run. Inside Copilot it’s available in VS Code, Visual Studio, Copilot CLI and on github.com.

The $0.30 cached input rate matters more than it looks. Agentic loops re-send a large stable prefix — repo context, tool definitions, instructions — on every step. A 10× cache discount is what makes long runs affordable.

The catch: it is billed at provider list pricing under usage-based billing, so it consumes budget rather than sitting inside a flat allowance. Watch spend before turning it loose on a fleet.

MAI-Code-1.1-Flash — the one most people will actually use

Microsoft’s own model is the strategic move here. It reaches Free and Student users through auto model selection, meaning it becomes the default coding model for the largest slice of Copilot’s user base without anyone picking it.

The differentiator is native image understanding. Paste a screenshot of a broken layout, a design mock, an error dialog or a whiteboard sketch and it works from the image directly. Among the fast tiers that is genuinely uncommon, and it maps to how developers actually debug UI.

The August 11 release also claims improvements to coding quality, instruction following, tool use and performance over MAI-Code-1-Flash. Treat those as vendor-reported until independent benchmarks land.

Enterprise note: admins may need to enable the model by policy before their organisation sees it.

The frontier models — keep them for the hard 10%

Claude Opus 5 ($5/$25, 1M context, 128K max output) remains the pick for long-horizon refactors where the deliverable is a large coherent change rather than an answer. GPT-5.6 Sol ($5/$30) is the strongest general reasoner. Both cost roughly 2× Kimi K3 per task, which is correct pricing for decisions that are expensive to get wrong and terrible pricing for autocomplete.

The Routing Policy That Works

Inline completion, small edits, anything with a screenshot → MAI-Code-1.1-Flash. Cheapest, fast, image-aware.
Multi-step agent tasks, tool-heavy runs → Kimi K3. Cache discount does the heavy lifting.
Cost-sensitive mid-tier work → Claude Sonnet 5 at $2/$10 — cheaper per task than K3 and permanently priced after Anthropic cancelled the September 2026 increase.
Architecture, gnarly debugging, large refactors → Claude Opus 5 or GPT-5.6 Sol.
Measure cost per completed task, not per token. A cheap model that needs three attempts is the expensive one.

What GitHub’s Strategy Signals

Adding a Chinese open-weight model and a Microsoft in-house model in the same week is not an accident. Copilot is positioning as a router, not a model — the value is the surface, the permissions, the enterprise controls and the IDE integration, with the model as a swappable component.

That’s the opposite of Cursor’s approach, which is vertically integrating everything from the model to code hosting. Model-choice platforms give you portability; vertically integrated ones give you speed. Both bets are live.

Last verified: August 18, 2026. Pricing from GitHub Copilot’s model pricing documentation.