Head-to-head comparison
Kimi K2 Thinking (Moonshot AI) and Kimi K3 (Moonshot AI) compared on the ModelCap Index, API price, context window, provider availability, weight access and every public benchmark board they share. Figures are the same ones shown on the live rankings; nothing here is a hidden score.
Snapshot as of 22 September 2026
As of 22 September 2026, Kimi K3 holds the stronger ModelCap Index position (#12 vs #47); Kimi K2 Thinking is 6× cheaper per output token ($2.50/1M vs $15.00/1M); and Kimi K3 offers the longer context window (1M vs 262K tokens).
Evidence: 2 public benchmark observations across 2 boards · 6 public benchmark observations across 6 boards. Prices are the lowest listed API offer per million tokens observed on the OpenRouter catalogue.
Scores are the sources' own published figures for each model's best evaluated configuration; ModelCap never re-runs a benchmark.
Want a different pairing? Open the interactive comparison tool to swap either model for any current ranked language model.
Kimi K3 ranks higher on the ModelCap Index as of 22 September 2026: #12 against #47. That is a capability ranking built from public benchmark evidence with published uncertainty; whether it is "better" for you also depends on price, context and where you can run it. Their published uncertainty intervals do not overlap, but the aggregate Index does not predict performance on every task; compare the relevant benchmark configurations before choosing.
Kimi K2 Thinking is cheaper on output tokens: $2.50/1M against $15.00/1M. Input tokens are $0.60/1M for Kimi K2 Thinking and $3.00/1M for Kimi K3. Prices are the lowest listed API offer ModelCap observed, in USD per million tokens. For 1,000 requests with 2,000 input and 500 output tokens each (2M input + 0.5M output), the listed-rate estimate is $2.45 for Kimi K2 Thinking versus $13.50 for Kimi K3. Kimi K2 Thinking costs 81.9% less in this scenario. This excludes caching, batch discounts, prompt-length tiers, tool charges and retries; verify the selected endpoint before budgeting.
Kimi K3 has the larger context window: 1M tokens against 262K.
Arena coding: Kimi K2 Thinking 1487, Kimi K3 1538 — Kimi K3 leads.
Kimi K2 Thinking: Restricted license. Kimi K3: Restricted license. Open weights mean the checkpoint can be downloaded and self-hosted under its licence; API-only models are available solely through hosted endpoints.
ModelCap currently lists 2 API providers for Kimi K2 Thinking and 17 for Kimi K3, from the OpenRouter catalogue snapshot the site serves; each model page lists the providers and their prices.
Every figure comes from the sealed ModelCap dataset published 22 September 2026; the page re-renders within a minute of each data refresh, and the ModelCap Index positions are the same ones shown on the live rankings.