On July 1, 2026, GitHub added a model to Copilot's picker that changes how you should think about the dropdown: Kimi K2.7 Code, from Moonshot AI — the first open-weight model ever offered as a selectable option there. It reached Pro, Pro+, and Max first; Business and Enterprise followed on July 7 behind an admin policy toggle. With it, the picker now spans five labs — OpenAI, Anthropic, Google, Microsoft, and Moonshot — behind one subscription. So the real question is no longer which vendor, it's when does the open model win.
Here's the whole decision in one screen, then the reasoning.
The three, side by side#
| Kimi K2.7 Code | Claude Sonnet 5 | GPT-5.6 Terra | |
|---|---|---|---|
| Weights | Open (MIT) | Closed | Closed |
| Input / 1M | $0.95 miss · $0.19 hit | ~$2 (promo) | $2.50 |
| Output / 1M | $4.00 | ~$10 (promo) | $15.00 |
| Self-hostable | Yes | No | No |
| Best for | high-volume, scoped coding | balanced quality | mature ecosystem |
By output price — the number that actually bills an agent — Kimi's token is 2.5–3.7x cheaper than the closed flagships. If you stop reading here you route everything to Kimi. Don't stop reading here.
Why it's this cheap: the MoE math#
Kimi K2.7 Code is a Mixture-of-Experts model: 1 trillion total parameters, but only about 32 billion activated per token. You get the knowledge capacity of a giant model while paying inference on a small one — that's the structural reason a frontier-adjacent coder can sell output at $4.00 per 1M tokens. The full weights are public on Hugging Face under an MIT license, and inside Copilot GitHub simply runs a hosted copy on Azure — so you select it like any other model, with no infra to manage.
Output is the number that bills an agent#
A chat app is input-heavy: long prompt, short answer, and cache discounts do most of the work. A coding agent is the opposite — it emits edits, tool calls, reasoning, and retries, turn after turn, so the output rate dominates the bill. That's exactly where Kimi's advantage compounds: $4.00 output against ~$10 for Sonnet 5's promo rate and $15 for Terra. On a token-heavy task — a codebase-wide refactor, a test-generation sweep, an agent loop — the gap is the difference between a $3 job and a $10 one.
The open-weight option means you stop routing on price alone. You route the easy 80% cheap and reserve the flagship for the 20% that genuinely needs it.
When the open model wins#
Pick Kimi K2.7 Code when the task is scoped and token-heavy:
- Codebase-wide edits — search-and-replace with judgment, mechanical migrations, dependency bumps across many files.
- Test scaffolding and boilerplate — high volume, low ambiguity, easy to verify.
- Agent loops that emit a lot of output per run, where the per-token saving multiplies across dozens of steps.
- Compliance and portability — an MIT-licensed, auditable, self-hostable weight is a hedge a closed endpoint can't offer. You get the convenience on Azure now with the option to leave with the weights later.
When to still pay for a flagship#
Keep Claude Sonnet 5, GPT-5.6 Sol, or Opus for the hardest reasoning and the most tangled multi-file debugging — the cases where a small quality gap wastes more of your time than the tokens cost. The point isn't that open beat closed; it's that you now have a rung below the flagships that's good enough for most work. The winning pattern is a ladder: default the easy majority to Kimi, detect when a step is going wrong, and escalate the hard minority up a tier. A gateway like LiteLLM makes that one line of config.
The week's real lesson: the model picker stopped being a taste test between vendors and became a cost-control dial. Open-weight didn't replace the flagships — it gave you somewhere cheaper to send the work that never needed one.
