Model comparison - Same provider, successor
Kimi K2.6 vs Kimi K3
Kimi K2.6 and Kimi K3 are both Moonshot AI models. On list price, Kimi K3 charges 3.2x Kimi K2.6's input price ($3 vs $0.95 per 1M tokens), and Kimi K3 charges 3.8x Kimi K2.6's output price ($15 vs $4 per 1M tokens). Kimi K2.6 has a 262,144-token context window and Kimi K3 1,048,576 tokens; Kimi K3 returns up to 1,048,576 output tokens per response, and no primary source states a max output for Kimi K2.6. Both publish open weights.
Last verified from each provider's pricing page. All sources
Side-by-side specs
| Spec | Kimi K2.6 | Kimi K3 |
|---|---|---|
| Provider | Moonshot AI | Moonshot AI |
| Input / 1M tokens | $0.95 | $3 |
| Output / 1M tokens | $4 | $15 |
| Cache read / 1M | $0.16 | $0.30 |
| Cache write / 1M | - | $3 |
| Batch input / 1M | $0.57 | - |
| Batch output / 1M | $2.40 | - |
| Context window | 262,144 | 1,048,576 |
| Max output | - | 1,048,576 |
| Weights | Open | Open |
| License | Modified MIT License (Hugging Face license tag: other) | Kimi K3 License (Hugging Face license tag: other) |
| API model id | kimi-k2.6 | kimi-k3 |
| AI Gateway id | moonshotai/kimi-k2.6 | moonshotai/kimi-k3 |
| Last verified |
Prices in USD per 1M tokens at each provider's standard first-party rate. A dash means the provider does not bill that item separately or no primary page states it.
What each costs in practice
The same three workloads every model page uses, worked from each model's standard list rates. Swap in your own token counts with the math on each model's page.
| Workload | Kimi K2.6 | Kimi K3 |
|---|---|---|
| One agent session1M input tokens and 100K output tokens, no caching. | $1.35 | $4.50 |
| A chat app10,000 requests a day at 2K input and 500 output tokens each. | $39.00$1,170.00 per 30-day month | $135.00$4,050.00 per 30-day month |
| The same session at a 90% cache hit90% of the input is read from the prompt cache. The one-time cache write premium is not included. | $0.639 | $2.07 |
Show the math
Kimi K2.6
- One agent sessionInput: 1M x $0.95 = $0.95Output: 0.1M x $4 = $0.40
- A chat appInput: 20M x $0.95 = $19.00Output: 5M x $4 = $20.00
- The same session at a 90% cache hitCached input: 0.9M x $0.16 = $0.144Uncached input: 0.1M x $0.95 = $0.095Output: 0.1M x $4 = $0.40
Kimi K3
- One agent sessionInput: 1M x $3 = $3.00Output: 0.1M x $15 = $1.50
- A chat appInput: 20M x $3 = $60.00Output: 5M x $15 = $75.00
- The same session at a 90% cache hitCached input: 0.9M x $0.30 = $0.27Uncached input: 0.1M x $3 = $0.30Output: 0.1M x $15 = $1.50
Benchmarks under the same setting
Only boards that list both models under the same run setting appear here, one row per board and setting, with each score exactly as the board shows it. Scores on different boards use different scales and are never combined.
| Board and setting | Kimi K2.6 | Kimi K3 |
|---|---|---|
| AIME 2026 (MathArena)MathArena harness, 4 runs per problem, thinking modeOfficial leaderboard - matharena.ai | 95.83% | 96.67% |
Set them up in your tools
Step-by-step guides whose official docs list Kimi K2.6 or Kimi K3 by model id.
- Use Kimi in Claude CodeLists Kimi K3 and Kimi K2.6
- Use Kimi in ClineLists Kimi K3 and Kimi K2.6
- Use Kimi in Codex CLILists Kimi K3
- Use Kimi in ContinueLists Kimi K3 and Kimi K2.6
- Use Kimi in GooseLists Kimi K3 and Kimi K2.6
- Use Kimi in OpenCodeLists Kimi K3 and Kimi K2.6
- Use Kimi in ZedLists Kimi K3 and Kimi K2.6
Model pages
FAQ
How do Kimi K2.6 and Kimi K3 compare on price?
On list price, Kimi K3 charges 3.2x Kimi K2.6's input price ($3 vs $0.95 per 1M tokens), and Kimi K3 charges 3.8x Kimi K2.6's output price ($15 vs $4 per 1M tokens). For one agent session of 1M input and 100K output tokens with no caching, that is $1.35 on Kimi K2.6 and $4.50 on Kimi K3. Prices are each provider's first-party list rates, verified October 17, 2026 and October 17, 2026.
What are the context windows of Kimi K2.6 and Kimi K3?
Kimi K2.6 has a 262,144-token context window and Kimi K3 1,048,576 tokens; Kimi K3 returns up to 1,048,576 output tokens per response, and no primary source states a max output for Kimi K2.6.
Are Kimi K2.6 and Kimi K3 open weights?
Kimi K2.6's weights are published at huggingface.co under the Modified MIT License (Hugging Face license tag: other). Kimi K3's weights are published at huggingface.co under the Kimi K3 License (Hugging Face license tag: other).
More comparisons
- Gemini 4 Argon vs Kimi K3Similar price, same tier
- GPT-6.1 Sol vs Kimi K3Similar price, same tier
- Claude Sonnet 5.5 vs Kimi K3Similar price, same tier
- Claude Opus 5.5 vs Kimi K3Similar price, same tier
- Grok 4.7 vs Kimi K3Similar price, same tier
- Kimi K3 vs Qwen3.8-2.4T-A95BSimilar price, same tier
- Kimi K3 vs Qwen3.8-MaxSimilar price, same tier
Sources
Kimi K2.6 verified from platform.kimi.ai; Kimi K3 verified from platform.kimi.ai. This page is as current as the older of the two: .
- https://platform.kimi.ai/docs/pricing/chat
- https://platform.kimi.ai/docs/guide/kimi-k2-6-quickstart
- https://platform.kimi.ai/docs/models.md
- https://platform.kimi.ai/docs/pricing/batch.md
- https://platform.kimi.ai/docs/platform-changelog.md
- https://www.kimi.ai/blog/kimi-k2-6
- https://huggingface.co/moonshotai/Kimi-K2.6
- https://platform.kimi.ai/docs/guide/kimi-k3-quickstart
- https://www.kimi.ai/blog/kimi-k3
- https://huggingface.co/moonshotai/Kimi-K3
Newsletter
Get model price changes as they ship
One email per week. New model releases, price cuts and what they mean for your bill.