← All models

Kimi K3

Language Available Full comparison ↗

moonshot/kimi-k3 · by Moonshot AI · mixture-of-experts

Pricing — 1 offering(s)

Moonshot AI

Input tokens

  • $3.00 / 1M tokens (input) Current 2026-07-16 → present

Cached input tokens

  • $0.30 / 1M tokens (cached input) Current 2026-07-16 → present

Cache write tokens (5-minute TTL)

  • $3.00 / 1M tokens (cache write) Current 2026-07-16 → present
    5-minute TTL (Moonshot's default). Writes with a 1-hour TTL cost $6.00/1M. A cache hit refreshes the entry's…

    5-minute TTL (Moonshot's default). Writes with a 1-hour TTL cost $6.00/1M. A cache hit refreshes the entry's lifetime with no further write charge.

Output tokens

  • $15.00 / 1M tokens (output) Current 2026-07-16 → present

Showing the active price and any recorded history. Full pricing history is available via the paid API — see API docs.

See how Kimi K3 fits into a cost-aware routing setup

See how →

Capability profile

How Kimi K3 rates across core capability dimensions, with the task-level evidence behind each rating.

coding Strong

SWE-bench Verified (500 human-verified GitHub issues; patch must pass the repo's own tests; no file-location hints): resolves 93.4% — fails roughly 1 issue in 15. Independent (Vals AI leaderboard, #3 of 72; mini-SWE-agent harness).

context window Strong

1,000,000-token context window per Moonshot's own model docs.

reasoning Strong
tool use Strong

AA-Briefcase Elo 1543 — the 2nd-highest score recorded on this independent agentic-knowledge-work benchmark, behind only Claude Fable 5 (1574), and well ahead of GPT-5.6 Sol max (1501), Claude Sonnet 5 max (1388), and Claude Opus 4.8 max (1347). A +727-point improvement over the previous-generation Kimi K2.6 (816) on the same benchmark. artificialanalysis.ai/articles/kimi-k3-agentic-knowledge-benchmark, published 2026-07-21, confirmed 2026-08-19.

instruction following Unknown
multilingual Unknown
speed Weak

37.7 output tokens/sec, rank #54/107 — Artificial Analysis describes it as "notably slow and somewhat verbose." Same source as reasoning, confirmed 2026-08-19.

cost efficiency Moderate

$0.84 cost per Artificial Analysis Intelligence Index task, rank #35/107. Same source, confirmed 2026-08-19.

Ratings are estimated — limited independent data is available for this model.

Coding benchmarks

View the full coding benchmark leaderboard →
Benchmark Score Harness Source
swe-bench-verified ↗ 93.4% 2026-07 mini-swe-agent 📊 leaderboard ↗
terminal-bench-2-1 88.3% 2026-07 kimi-code 🏢 vendor ↗
terminal-bench-4-0 19.7% 2026-09 mini-swe-agent 📊 leaderboard ↗

Operator guidance

Moonshot's current flagship (superseded K2.6, 2026-07-16). Independently confirmed strong on coding (SWE-bench Verified) and now also on reasoning and tool-use (Artificial Analysis Intelligence Index and AA-Briefcase, both #1-2 in class) — the strongest-evidenced entry in the Kimi family. Notably slow output (37.7 tok/s) is the main tradeoff. Drop to Kimi K2.6 for the prior-generation, lower-cost option, or K2.7 Code specifically for coding-focused workloads at a much lower price.

Use cases

Limitations

Citations