Kimi K2.7 Code is generally available in GitHub Copilot

Original Article ↗ Hacker News Discussion ↗

GitHub Copilot’s integration of the Kimi K2.7 Code model is welcomed as a cheaper, open‑weight alternative to US “frontier” models, especially for enterprises that want access to Chinese models from a trusted Western provider. However, many users say Copilot’s recent shift to token-based pricing has made it dramatically more expensive and less attractive than rivals like Claude Code, Cursor, OpenCode, or even running Qwen and other models locally. The thread also highlights how model quality depends heavily on the surrounding “harness,” and how frustration with vendor lock‑in and sudden pricing changes is pushing developers toward local or multi-provider setups.

Custom / BYO models in Copilot and VS Code

  • Commenters say Copilot already supports custom/BYOM models in some contexts (CLI, Copilot app, VS Code), but integration is poorly documented and unstable.
  • VS Code’s “bring your own key” (BYOK) support for Copilot Chat is described as the more reliable path; people report successfully wiring in third‑party models via this route.
  • Some users report failures trying to add certain external LLMs directly as Copilot custom models, but success when going through VS Code BYOK instead.

Kimi K2.7-Code, other models, and hosting

  • Many see Kimi K2.7-Code as a welcome non‑US option, especially for enterprises that can’t directly use Chinese providers but can use GitHub/Microsoft.
  • Benchmarks are cited that put it near “Sonnet 4.6”–class performance, but others are skeptical, reporting worse real‑world results than Claude, possibly due to quantization or hosting differences.
  • GitHub docs say Moonshot models (Kimi) are hosted on US‑based Azure AI Foundry; prompts aren’t sent to the original model developer.
  • Commenters expect GLM 5.2 and DeepSeek V4 to be added next, noting they’re already on Azure.

Pricing, multipliers, and user migration

  • The recent Copilot shift to strict token billing and model multipliers is widely described as a “rug pull” and “insane” price hike.
  • Many individual users and teams say their $10–$20 monthly budgets are now exhausted in days, especially using Claude Opus–class models.
  • Several orgs report cancelling Copilot and moving to Claude Code, Cursor, liteLLM, OpenCode, or direct API use; some still find Copilot cheaper at enterprise scale due to bundled credits and negotiated discounts.
  • Confusion around annual/legacy plans and model multipliers is noted; some features/models appear unavailable or inconsistently documented.

Harness quality and experience

  • Multiple comments claim Claude (especially Claude Code) outperforms the same models via Copilot, attributing this to better “harnesses” (system prompts, tools, routing).
  • Others like Copilot CLI and Agent Client Protocol integration, praising multi‑model orchestration and sub‑agent routing.
  • Zed, Reasonix, and other harnesses are discussed; tool‑calling reliability and caching strategies dramatically affect perceived model quality and cost.

Cloud, compliance, and alternatives

  • Some worry about CLOUD Act exposure and doubt GDPR compliance when using US clouds; others argue most European infra is already on those clouds, so it’s a moot point.
  • Open weight hosting via Copilot, Cloudflare, Synthetic, Fireworks, etc. is seen as giving enterprises cheaper or more flexible options.
  • Several commenters advocate moving to local models (Qwen, Gemma, GLM, etc.), valuing cost control, stability, and extensibility over absolute frontier performance.