Kimi ships K3-256k: same output as its 1M model at half the quota

Kimi Code, the AI coding assistant from Moonshot, has added a new model ID called k3-256k. It caps the context window at 256k tokens but, within that window, produces results identical to the full 1M-context k3 while consuming roughly half the quota. Kimi pitches it as the default for everyday work — Q&A, code completion, routine feature development, and single- or small-file edits. It does not accept video input.

The more interesting details are operational. Context caches are tied to a specific model, so switching model IDs — or even changing reasoning-effort levels — invalidates the cache and forces a re-prefill, which makes usage spike immediately after a switch; Kimi’s fix is simply to start a fresh session. Downgrading from 1M to 256k mid-session triggers a client-side compaction in tools like the Kimi CLI and Claude Code when the context already exceeds 256k, and any session containing video files will fail the switch outright.

Access is gated by subscription tier: plans below Moderato can’t call k3 at all, 1M context requires Allegretto or higher, and requests that exceed a plan’s entitlements return a 401. Two silent-failure traps stand out for anyone wiring this into third-party tools over Kimi’s OpenAI- or Anthropic-compatible endpoints — mistyping the HighSpeed model ID quietly falls back to standard speed with no error, and turning off ‘thinking’ silently reroutes K3/K2.7 requests down to the older K2.6 model.