GitHub added Kimi K3 to Copilot on August 6 — and it quietly changed the model picker for millions of developers. Moonshot AI’s 2.8-trillion-parameter open-weight model is now generally available across all Copilot plans, priced at $3 per million input tokens. That undercuts Claude Opus 5 ($5/M) and most GPT-5 variants on a straight per-token comparison. It is also off by default for every Business and Enterprise organization, which is the part worth paying attention to.
How to Enable Kimi K3 #
For individual plan holders — Pro, Pro+, and Max — the process takes about ten seconds. Open Copilot Chat, click the model picker, select Kimi K3, and you are done. The model is available in VS Code, Visual Studio, JetBrains, Xcode, Eclipse, Copilot CLI, GitHub Mobile, and github.com.
For Business and Enterprise accounts, nothing happens automatically. A plan administrator must navigate to the organization’s Copilot policy settings and explicitly enable the Kimi K3 model policy. Until that policy is active, no one in the organization can select the model. GitHub left this off by default deliberately — open-weight models routed through a third-party inference provider require a compliance review before they reach developer workstations.
What Kimi K3 Is Actually Good At #
The benchmarks are strong but worth parsing carefully. On SWE-bench Verified, Kimi K3 scores 93.40% — behind Claude Opus 5 (97.00%) and GPT-5.6 Sol (96.20%), but a meaningful step up from Claude Opus 4.8 (88.60%). That gap closes in agentic workloads. K3 takes the top spot on LMArena’s Frontend Code Arena across six of seven domains, ranks first on Program Bench and SWE Marathon, and finishes second on Terminal-Bench 2.1.
The pattern is consistent: K3 wins in iterative, multi-step coding loops. It loses ground on single-pass deep file analysis, where Claude Fable 5 and GPT-5.6 Sol still lead. If your Copilot workflow involves long agentic sessions — generating boilerplate, iterating on front-end components, running repeated refinement passes — K3’s performance profile is a genuine fit. If you need one-shot architectural review of a large codebase, the top proprietary models remain the better call.
The Pricing Case #
Kimi K3 is priced at $3 per million input tokens, $15 per million output tokens, and $0.30 per million cached input tokens. Relative to Claude Opus 5 at $5/$25 input/output, the savings compound quickly in agentic contexts where models generate substantial output per session. For teams running automated coding agents at any meaningful scale, routing qualifying workloads to K3 could cut inference costs significantly without sacrificing competitive benchmark performance.
The Enterprise Caveat: Open Weight Is Not Self-Hosted #
This is the detail that gets glossed over in the excitement. Kimi K3’s weights are public, but the Copilot integration routes your code through Fireworks AI — not GitHub’s own infrastructure. Fireworks is the inference provider GitHub selected for K3. They offer Zero Data Retention by default and a US-only serverless endpoint for regulated industries, but it remains a third party in the data chain.
For most development teams, this is a non-issue. For financial services, healthcare, and other compliance-heavy organizations, it requires a policy review before the administrator enables the model. Fireworks has published documentation on their data residency and retention posture for this audience. The US-only endpoint addresses data sovereignty requirements for many regulated use cases. But the core point stands: enabling Kimi K3 in an enterprise Copilot account is an infrastructure decision, not just a model preference.
The Bigger Picture #
Kimi K3 follows Kimi K2.7 Code, which went GA in GitHub Copilot on July 1. Moonshot AI has shipped two frontier-class models into production developer tooling within six weeks. The pattern is significant: a Chinese AI lab is now delivering open-weight models inside a Microsoft product to enterprise development teams worldwide. Forbes called it a convergence toward open-weight models — and the Copilot integration makes that concrete.
For most Copilot users, Kimi K3 is worth enabling and testing on front-end and agentic workflows. The benchmarks support it, the pricing makes experimentation cheap, and the model picker makes switching frictionless. For enterprise admins, the right move is a policy review before you flip the switch — not because K3 is unsafe, but because any new model routing code through a new infrastructure provider warrants that level of scrutiny.