Kimi K3 is Moonshot AI's 2.8-trillion-parameter open-weights model with a 1M-token context window. On the Kimi Open Platform it costs $3.00 per 1M input tokens and $15.00 per 1M output tokens; Moonshot now calls it its previous-generation flagship, after the release of Kimi K3.1 Pro.
Kimi K3 costs $3.00 input / $15.00 output per 1M tokens, with cached input at $0.30. Batch jobs are billed at 60% of standard prices, and because the full weights are public you can also self-host it or buy it from other inference providers.
| Kimi K3 | |
|---|---|
| Provider | Moonshot AI |
| Tier | Flagship |
| Input / 1M tokens | $3.00 |
| Output / 1M tokens | $15.00 |
| Cache hit / 1M | $0.30 |
| Cache write / 1M | $3.00 |
| Context window | 1M |
| Notes | Previous-generation flagship, superseded by Kimi K3.1 Pro but still sold; open weights; cache write $3 (5-min TTL) or $6 (1-h TTL) |
| Model | Support chatbot 10,000 conversations × (2,000 input + 400 output tokens) | Document Q&A 1,000 questions × (30,000 input + 800 output tokens) | Coding agent session 4M input tokens (90% cache hits) + 150,000 output tokens |
|---|---|---|---|
| Kimi K3 | $120 | $102 | $4.53 |
| Grok 4.7 | $64.00 | $64.80 | $3.50 |
| Qwen3.8-Max | $64.00 | $64.80 | $2.31 |
| Gemini 4 Argon | $80.00 | $68.00 | $2.66 |
Standard rates per 1M tokens from data on this site. The coding-agent row assumes 90% of input is read from the prompt cache (full input price where a model has no cache rate). Cache-write fees, batch discounts and long-context surcharges listed in the notes are not included. Updated November 2, 2026.
| Model | Tier | Input / 1M | Output / 1M | Cache hit / 1M | Context | Notes |
|---|---|---|---|---|---|---|
| Kimi K3.1 Pro | Flagship | $2.00 | $10.00 | $0.20 | 1M | Moonshot's current flagship, a further post-trained K3 at a lower price; cache write $2 (5-min TTL) or $4 (1-h TTL); Highspeed variant $4/$20 |
| Kimi K3 | Flagship | $3.00 | $15.00 | $0.30 | 1M | Previous-generation flagship, superseded by Kimi K3.1 Pro but still sold; open weights; cache write $3 (5-min TTL) or $6 (1-h TTL) |
| Model | Tier | Input / 1M | Output / 1M | Cache hit / 1M | Context | Notes |
|---|---|---|---|---|---|---|
| Grok 4.7 | Flagship | $2.00 | $6.00 | $0.50 | 500K | New flagship (Sep 21, 2026); larger base model, replaces 4.6 at the same price; ≥200K tokens: $4/$12, cache $1.00 |
| Qwen3.8-Max | Flagship | $2.00 | $6.00 | $0.17 | 1M | New Alibaba flagship (Aug 3, 2026); 2.4T MoE |
| Gemini 4 Argon | Flagship | $2.00 | $10.00 | $0.10 | 1M | New top frontier model, generally available Oct 13, 2026 for AI Pro/Ultra subscribers and via the API (model id gemini-4-argon); long-decode generation up to 1M tokens; intro price thru Nov 30, 2026, then $4/$20 (cache read $0.20); not available on the free tier |
| Gemini 3.1 Pro | Flagship | $2.00 | $12.00 | $0.20 | 1M | Frontier multimodal (>200K: $4/$18, cache $0.40) |
| GPT-5.6 Sol | Flagship | $4.00 | $20.00 | $0.40 | 1M | Cut 20% Aug 21 (promo thru Nov 21); >272K tokens: $8/$30, cache $0.80 |
Kimi caches repeated request prefixes automatically; you don't create caches yourself. Hits are billed at $0.30, but for K3-series models each cache write is a separate one-time charge, priced by TTL: 5m (the default) or 1h. The tables on this page don't include those write fees. Pick the TTL from your request rhythm: rapid back-and-forth agents are fine on five minutes, while workloads with gaps longer than that should pass prompt_cache_options.ttl to use the one-hour tier and avoid paying for the same write repeatedly. The cache is stored in blocks, and a partial block that doesn't fill a whole block is billed as a cache miss.
K3 is unlocked after a small minimum first top-up, and your cumulative top-up total sets your account tier and rate limits (concurrency, RPM, TPM and TPD); vouchers don't count toward that total. Billing is pay-as-you-go with no subscription. For offline jobs such as evaluations, labeling or bulk summarization, the Batch API bills kimi-k3 at 60% of standard prices.
The API works with the OpenAI SDK. K3 always runs with thinking enabled; you control depth with reasoning_effort (low, high or max, default max), and lower effort is the main lever for cutting output tokens at $15.00 per 1M. It accepts image input, but only as base64 or an uploaded file reference, not public image URLs. In multi-turn and tool-calling flows, send back the complete assistant message, not only its content. Moonshot now presents K3.1 Pro, a post-trained successor, as its flagship, and its price list puts K3.1 Pro below K3 on both input and output, so compare the two before committing to K3 for a new project.
$3.00 per 1M input tokens, $15.00 per 1M output tokens and $0.30 per 1M cached input tokens, plus a separate cache-write charge that depends on the TTL you choose.
Yes. Moonshot has released the full Kimi K3 weights, and describes it as the first open-source model at the 2.8-trillion-parameter scale.
Yes. Batch inference for kimi-k3 is billed at 60% of the standard price.
Yes. Perplexity's Agent API lists perplexity/kimi-k3 at the same per-token input, output and cache-read rates shown here.
K3.1 Pro is a post-trained version of K3 with more post-training compute, Moonshot now presents it as its flagship, and it lists lower input and output prices than K3. Both have a 1M-token context window and support batch inference at 60% of standard prices.
How we choose: prices come from each vendor's official pricing page and are re-checked every week. We take no affiliate fees, so rankings aren't for sale.