Kimi K3 API Pricing

Moonshot's pricing page lists $3.00/M cache-miss input, $0.30/M cache-hit input and $15.00/M output, for a 1,048,576-token context window. Under AI//COST's 92/8 agentic blend with 82% cache hits, the effective planning figure is $1.92/M. Prices exclude local taxes, which the vendor calculates at checkout.

Kimi K3.1 Pro, unless you have a reason to pin K3. Moonshot describes K3.1 Pro as a post-trained K3, it is a third cheaper on every rate - $2.00/M input and $10.00/M output against $3.00 and $15.00 - and Moonshot's model list now calls K3 its previous-generation flagship. K3 has no deprecation date.

Yes. When this page first went up on July 16, 2026 the rates were live on platform.kimi.ai with no launch post behind them, so we published them as vendor-listed but pre-announcement. Moonshot has since published the launch post on kimi.com/blog, put a launch banner across the API docs, and made K3 the flagship on moonshot.ai, which describes it as a 2.8T-parameter natively multimodal model. Re-verified on July 28, 2026: the prices did not change with the announcement.

K3 costs roughly 3× more on input ($3/M vs $0.95/M) and 3.75× more on output ($15/M vs $4/M). On our agentic blend that is $1.92/M effective against $0.60/M for K2.6. If your workload does not need the 1M context window or K3's reasoning depth, Kimi K2.6 remains substantially cheaper.

Automatically, with a price on the write. Kimi writes eligible request prefixes to its cache with a 5-minute TTL unless you ask for 1 hour (prompt_cache_options on Chat Completions and Responses, top-level cache_control on the Anthropic-compatible Messages API, which only writes when you set it). Since September 2026 the card bills the write as its own line - $3.00/M for 5 minutes, $6.00/M for 1 hour - and Moonshot says that is not a price increase: the write replaces the input charge on the part that missed, so a default request costs what it did. Cache hits are billed at $0.30/M instead of $3/M, a 90% discount, and the calculator's default agentic blend assumes an 82% cache-hit rate. The 1-hour TTL pays for itself after two hits within the hour.

AI//COST uses the same 92/8 agentic blend everywhere: 92% of tokens are input, 82% of which are served from cache. For Kimi K3 that gives a blended input of $0.786/M and an effective blended cost of $1.92/M. The same formula is applied to every model in the shelf above, so the numbers are directly comparable.

Yes, indirectly. K3 always has thinking mode enabled and currently only supports reasoning_effort: max, so reasoning tokens are billed as output at $15/M. There is no cheaper non-thinking variant of K3, unlike the K2 family. Output-heavy workloads will feel this more than input-heavy ones.

Kimi K3 is documented with native visual understanding and accepts text, image, and video input. The pricing page bills consumed tokens, so extracted document, image, or video context should be treated as model input once passed into a request.

The browser widget uses a moonshot-tokenizer-estimate chars-per-token estimate for English text. Actual billing comes from Kimi API usage fields and can differ for Chinese, code, images, video, or mixed-language prompts.