Kimi K3
TL;DR
Open-sourceMoonshot's flagship — long-horizon coding and end-to-end knowledge work with a 1M-token context.
Available on
Routed across multiple upstream providers; price matches the Command Code pricing page.
Switch with
Pick Kimi K3 from the selector.
Input
$3
per M tokens
Output
$15
per M tokens
Cache read
$0.30
per M tokens
Kimi K3 in Command Code
Kimi K3 is Moonshot's flagship model for long-horizon coding and end-to-end knowledge work, with a 1M-token context window and vision input. It always thinks — Moonshot calls this Preserved Thinking — so there is no reasoning-effort knob to tune; K3 runs at maximum depth on every request. In Command Code it sits at the top of the Kimi line, above the K2.7 Code models.
K3 vs the Command Code lineup
Pricing for K3 alongside the most relevant peers. K3 is not yet on the Intelligence Index; benchmarked peers show their scores.
| Model | Intelligence | Speed | Input $/M | Output $/M |
|---|---|---|---|---|
| Kimi K3 | Not yet scored | — | $3.00 | $15.00 |
| Kimi K2.7 Code | Not yet scored | — | $0.95 | $4.00 |
| Kimi K2.6 | 54 | ~40 tok/s | $0.95 | $4.00 |
| GLM-5.2 | Not yet scored | — | $1.40 | $4.40 |
| Claude Sonnet 4.6 | 52 | ~62 tok/s | $3.00 | $15.00 |
What K3 is best for
Whole-repository work that needs the full 1M-token window, long-horizon agent runs where the model must hold context across many turns, and knowledge work that mixes code, documents, and screenshots (vision).
When to switch away from K3
Switch to Kimi K2.7 Code
Roughly a third of K3 input cost and a quarter of its output cost ($0.95/$4.00). Pick it when 256K of context is enough and you do not need K3 depth on every request.
Switch to Kimi K2.7 Code HighSpeed
When latency matters more than per-token cost — the same K2.7 Code model at roughly 2× throughput ($1.90/$8.00).
Switch to GLM-5.2
Another 1M-context option at less than half the input price ($1.40/$4.40), with a high/max reasoning-effort ladder when you want to dial depth down.
In Command Code: caching and taste-1
Open-source models are routed across multiple upstream providers for high availability. The price you see is the mean per-provider rate; the Usage page reflects what was actually charged.
Where supported by the upstream, prompt caching is on by default — cache reads are billed at $0.30 per million tokens versus $3 for fresh input.
taste-1 sits between the model and the agent loop, rewriting and reranking candidate edits to match your codebase conventions.
Plan availability
Open-source model. Available on every plan, including Go ($1/mo). Routed across multiple upstream providers; listed price matches the Command Code pricing page.
All Command Code models, ranked by quality and speed
Quality is the Intelligence Index — an aggregate score across reasoning, math, coding, and knowledge evaluations. Speed is reported output tokens per second. Models without a published score are noted.
| Model | Tier | Intelligence Index | Output speed |
|---|---|---|---|
| GPT-5.5 | Premium | 60 | ~65 tok/s |
| Claude Opus 4.7 | Premium | 57 | ~49 tok/s |
| GPT-5.4 | Premium | 57 | ~84 tok/s |
| GPT-5.3 Codex | Premium | 54 | ~72 tok/s |
| Kimi K2.6 | Open-source | 54 | ~40 tok/s |
| Claude Sonnet 4.6 | Premium | 52 | ~62 tok/s |
| DeepSeek V4 Pro | Open-source | 52 | ~35 tok/s |
| GLM-5 | Open-source | 50 | ~61 tok/s |
| GPT-5.4 Mini | Premium | 49 | ~164 tok/s |
| DeepSeek V4 Flash | Open-source | 47 | ~82 tok/s |
| Claude Haiku 4.5 | Premium | 37 | ~97 tok/s |
| Kimi K2.5 | Open-source | 37 | ~35 tok/s |
| Claude Sonnet 5 | Premium | Not yet scored | — |
| Claude Opus 4.8 | Premium | Not yet scored | — |
| Claude Opus 4.6 | Premium | Not yet scored | — |
| GPT-5.6 Sol | Premium | Not yet scored | — |
| GPT-5.6 Terra | Premium | Not yet scored | — |
| GPT-5.6 Luna | Premium | Not yet scored | — |
| Muse Spark 1.1 | Premium | Not yet scored | — |
| Kimi K3 | Open-source | Not yet scored | — |
| Tencent Hy3 | Open-source | Not yet scored | — |
| MiniMax M2.5 | Open-source | Not yet scored | — |
| Inkling | Open-source | Not yet scored | — |
Switching models with /model
In an interactive Command Code session, run /model to open the model selector. Pick the model you want and it applies to this session and to future sessions until you change it again. Premium models require Pro or higher; open-source models are available on every plan, including Go.
cmd # start an interactive session
/model # open the selector and pick a modelPlans and pricing
Command Code is a subscription with model usage at API rates. Each plan ships with monthly LLM credits. Credits roll over and never expire. Auto top-up keeps you running if you go over.
| Plan | Price/mo | LLM credits | Models |
|---|---|---|---|
| Go | $1 | $10 | Open-source only |
| Pro | $15 | $30 | Open-source + premium |
| Max 10× | $100 | $150 | Open-source + premium |
| Max 20× | $200 | $300 | Open-source + premium |
| Teams | $40 / seat | Pooled | Open-source + premium |
| Enterprise | Custom | Custom | Custom pool, SSO, audit logs |
Frequently asked questions
Kimi K3 or Kimi K2.7 Code?
K3 is the flagship: a 1M-token context window (vs 256K) and always-on maximum-depth thinking, at $3.00/$15.00. K2.7 Code is far cheaper at $0.95/$4.00 and enough for most coding runs. Switch with /model.
Can I turn K3 thinking down?
No. K3 always thinks with Preserved Thinking, and its reasoning effort only supports max, so Command Code shows no effort selector for it. Use GLM-5.2 if you want a 1M-context model with a reasoning-effort ladder.
Why no Intelligence Index for K3?
Public aggregate benchmarks have not yet been published for K3 in the current Intelligence Index format. The model is available and routed normally.
Which Command Code model should I use?
For open models, Kimi, DeepSeek, Qwen, MiMo are all strong picks. For closed models, Claude Sonnet 5 is the recommended default — the best combination of speed and intelligence, and a drop-in upgrade from Sonnet 4.6. Switch to Claude Opus 4.8 (the newest Anthropic flagship) for the most capable long-horizon agentic coding, GPT-5.5 (Intelligence Index 60) for the absolute hardest reasoning, or Claude Opus 4.7 / GPT-5.4 (both 57) for top-tier work at lower cost. For fast lookups, Claude Haiku 4.5 or GPT-5.4 Mini. For open-source, Kimi K2.6 leads the open-weights tier (Intelligence Index 54).
Can I mix Kimi K3 with other models in a workflow?
Yes. Switch per session using /model. Common pattern: keep Sonnet 5 as the default and switch up to Opus 4.8 or down to Haiku 4.5 as the task calls for it.
Are open-source model prices fixed?
Open-source models are routed across multiple upstream providers for high availability. The price listed for each is the mean per-provider rate. Actual cost on a given request may vary slightly. The Usage page reflects the price charged.
Is Command Code free to try?
The Go plan starts at $1/mo with $10 in LLM credits. It covers open-source models only. Pro at $15/mo unlocks premium models with $30 in LLM credits.
Does Command Code train on my code?
No. Command Code does not train on your code or store your code snippets. taste-1 data is stored locally in your project directory.
Where can I track my usage?
The Usage page in Studio shows per-request cost, token counts, and which model ran. Settings > Billing lets you change plans, buy credits, or enable auto top-up.
Does Command Code replace my editor?
No. Command Code is editor-agnostic — it runs as a CLI and works alongside any editor (Cursor, VS Code, Zed, JetBrains, Neovim, etc.).
Ship code that matches your taste
Command Code is the AI coding agent that continuously learns your taste. Start for $1.