Kimi K3 API

Price check: official, OpenRouter and Wokey

USD per million tokens. Official and OpenRouter rates were checked by hand on 2026-09-23, excluding tax; the Wokey rate is read from the live price list.

The cheapest OpenRouter host is Sail Research's fp4-quantized deployment ($1.4989 input / $10.758 output). Quantized output can differ from the official weights, so it is not in the table.

Checked 2026-09-23 · Official pricing page · OpenRouter endpoint list

Use it in Claude Code

This model supports /v1/messages on Wokey. Set these variables, then start claude.

export ANTHROPIC_BASE_URL="https://api.wokey.ai"
export ANTHROPIC_AUTH_TOKEN="$WOKEY_API_KEY"
export ANTHROPIC_MODEL="kimi-k3"

Kimi K3 API quickstart

Get a Wokey API key and follow these three steps to send your first request.

1. Get an API key

   Register or sign in to Wokey, open the API console, and create and copy a key in API key management. Check your available balance before calling.

2. Configure your client

   Use the base URL below in your OpenAI-compatible client, enter your Wokey API key, and select kimi-k3 as the model.

3. Send your first request

   Replace YOUR_WOKEY_API_KEY in the example with your own key and run it in a terminal. Read the answer in choices, then review usage and charges in the API console.

Get a Wokey API key · Read API docs

Client base URL

https://api.wokey.ai/v1

Model ID

kimi-k3

Full request URL

https://api.wokey.ai/v1/chat/completions

The client base URL includes /v1. The curl example uses the full request URL.

Minimal request

export WOKEY_API_KEY="YOUR_WOKEY_API_KEY"

curl https://api.wokey.ai/v1/chat/completions \
  -H "Authorization: Bearer $WOKEY_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "kimi-k3",
    "messages": [{"role": "user", "content": "Hello, Kimi!"}],
    "max_tokens": 1024
  }'

Kimi K3 at a glance

Release date

2026-07-17, on the Kimi web app, the Kimi Work desktop app, Kimi Code and the Kimi API platform on day one; the full weights went public on Hugging Face on 2026-07-27

Positioning

Moonshot says K3 still trails Claude Fable 5 and GPT-5.6 Sol overall but outperformed every other model in its evaluations, and recommends Kimi Code CLI as its agent framework

Model and weights

A 2.8T-parameter MoE with 104B active, routing each token to 16 of 896 experts, built on Kimi Delta Attention and Attention Residuals; weights at moonshotai/Kimi-K3, quantized to MXFP4 during training, about 1.56 TB in total, under the Kimi K3 License

Context and output

1,048,576-token context; max_completion_tokens defaults to 131,072 and goes up to 1,048,576; native text, image and video input, text output

Thinking and effort

Thinking is always on and responses carry reasoning_content; reasoning_effort accepts low / high / max with max as the default (only max existed at launch); temperature, top_p and the other sampling parameters are fixed and should be omitted

Official price

$3 input / $15 output per million tokens and $0.30 for cache hits; no context-length tiers

API and tools

The Kimi platform offers OpenAI- and Anthropic-compatible APIs with tool calling (tool_choice can be required), json_schema structured output and context caching; Wokey serves /v1/chat/completions and /v1/messages

Checked 2026-10-02 · Sources: Moonshot · Hugging Face · K3 quickstart · Pricing · Artificial Analysis

Official benchmarks: Kimi K3 vs Fable 5 vs GPT-5.6 Sol vs Opus 4.8 vs GLM-5.2

From Moonshot’s model card on Hugging Face; these are vendor-reported results, and “—” means no score was published. The comparison models are those current at K3’s launch (2026-07), so Claude Opus 5.5 and GLM-5.3, both released later, are not in the table; see “Which to use” below for those comparisons.

Every K3 score uses reasoning_effort=max and temperature=1.0. The Fable 5 column ran with fallback, the card’s GPT-5.5 column is left out, and the HLE-Full row is the with-tools score. On DeepSWE, Terminal-Bench 2.1 and ProgramBench K3 ran in the Kimi Code harness, while most other scores come from public leaderboards or each model’s best harness, so the conditions differ: on the official DeepSWE leaderboard’s mini-SWE-agent harness K3 scores 67.3%. The SWE-Marathon row uses an H20-recalibrated task set from before the v1.1 release; Z.ai’s table gives K3 48.1% on v1.1.

Which to use: Kimi K3, Claude Opus 5.5 or GLM-5.3

vs Claude Opus 5.5

Opus 5.5 shipped on 2026-09-22, two months after K3. On Artificial Analysis’ Intelligence Index v4.3.2 Opus 5.5 scores 58 at max and K3 44, a clear gap. K3 lists at $3 / $15, only 25% below Opus 5.5’s $4 / $20, but Opus 5.5 used 260M output tokens across the index against K3’s 160M, so a task costs $5.98 versus $2.00 at official prices, about a third for K3. Opus 5.5 outputs about 92 tokens/s and K3 about 34. Pick Opus 5.5 for the hardest coding, long-horizon agents and work that has to be right the first time; pick K3 for high volume, tight budgets, video input or self-hosting.

vs GLM-5.3

The two are level on Artificial Analysis’ Intelligence Index: GLM-5.3 scores 45 at max and K3 44, at almost the same cost per task ($2.01 vs $2.00). GLM-5.3 lists at $1.4 / $4.4, less than half of K3, but writes longer outputs (210M tokens across the index versus 160M) and runs about twice as fast (about 66 vs 34 tokens/s). Z.ai’s GLM-5.3 launch page lists both: they are close on DeepSWE v1.1 (66.9% vs 67.5%), GLM-5.3 leads on Terminal Bench 3.0 (28.3% vs 17.4%), and K3 leads on SWE-Marathon v1.1 (48.1% vs 42.5%) and Toolathlon (76.5% vs 73.0%). Either works for text-only coding; pick K3 for image or video input, since GLM-5.3 is text-only; for self-hosting, GLM-5.3 (753B total parameters) is far easier than K3 (2.8T).

Choosing effort

K3 has only low, high and max, with max as the default, and thinking cannot be turned off. Artificial Analysis measured max at 44 for $2.00 per task and low at 30 for $1.15: low is about 40% cheaper per task but drops 14 points. Use the default max for coding and agents; there is no third-party data for high yet, so compare it with max on your own tasks; keep low for simple Q&A and format conversion. Reasoning tokens bill at the output rate, so higher effort means more output.

Kimi effort docs

Third-party and community reviews

• Artificial Analysis: On Intelligence Index v4.3.2 it scores 44 at max and 30 at low; it used 160M output tokens across the index for $2.00 per task, outputs about 34 tokens/s and takes about 4.5 seconds to first token. On launch day, on the index version of the time, it scored 57 and ranked third, level with Claude Opus 4.8 and GPT-5.5 and behind Fable 5 and GPT-5.6 Sol, while its hallucination rate rose from K2.6’s 39% to 51%. Every Wokey meter is 30% of the official rate, so the same task comes to about $0.60 at max and $0.34 at low.
• Fireworks: Fireworks, which also serves K3, ran it against Claude Fable 5 in the same agent loop on about 1,030 agentic tasks. On SWE tasks K3 passed 92.4% and Fable 5 92.6%, essentially level, but K3 took about 55 turns and 1.3M tokens per task against Fable 5’s 21 turns and 130K tokens.
• Hacker News (2,107-point thread): Price drew the most discussion: many found $3 / $15 steep for an open-weights model, while others said it is fair if K3 really trails only Fable and Sol. Next came thinking length: some complained it thinks too long and repeats itself; Simon Willison’s pelican test produced 16,658 output tokens, 13,241 of them reasoning, for about $0.25; another commenter measured about 60% of K2.6’s reasoning tokens.

Using K3 in Claude Code, and request rules to check

These rules come from Kimi’s K3 quickstart and from issues Wokey has debugged in production. Check them before switching from K2.x to kimi-k3 or wiring it up for the first time.

• Thinking cannot be turned off, and responses always carry reasoning_content. When migrating from K2.x, replace the old thinking parameter with the top-level reasoning_effort (low / high / max).
• In multi-turn chats and tool calls, pass back the complete assistant message the API returned, including reasoning_content and tool_calls; sending back only content drops the earlier reasoning.
• temperature, top_p, n, presence_penalty and frequency_penalty are fixed at 1.0, 0.95, 1, 0 and 0, and Kimi asks you to omit them from requests.
• Images cannot be public URLs: put a base64 data URL such as data:image/png;base64,... in image_url, because an https image link returns 400 unsupported image url; video also goes in as base64. Kimi’s upload-then-reference flow with ms://<file_id> is not available through Wokey yet.
• max_completion_tokens defaults to 131,072 and goes up to 1,048,576. Reasoning tokens bill as output and the default max effort writes long outputs, so leave room in output budgets.
• With a stop sequence set, K3 can be cut off while still reasoning and return finish_reason stop with empty content. That comes from upstream and Wokey does not remove text, so test stop-dependent flows on your own tasks first.
• Using it in Claude Code: kimi-k3 on Wokey supports /v1/messages, so set ANTHROPIC_BASE_URL to https://api.wokey.ai, ANTHROPIC_AUTH_TOKEN to your Wokey API key and ANTHROPIC_MODEL to kimi-k3; it costs $0.90 input / $4.50 output per million tokens. Moonshot recommends Kimi Code CLI as the agent framework, and OpenCode and other OpenAI-compatible clients use /v1/chat/completions.

Request body with an image (/v1/chat/completions)

{
  "model": "kimi-k3",
  "reasoning_effort": "high",
  "messages": [{
    "role": "user",
    "content": [
      { "type": "image_url", "image_url": { "url": "data:image/png;base64,iVBORw0KGgo..." } },
      { "type": "text", "text": "Find the layout bug in this screenshot" }
    ]
  }]
}

Catalog pricing snapshot

Prices are per 1M tokens; request-time pricing controls settlement.

Official price source: https://platform.kimi.ai/docs/pricing/chat-k3

Model capabilities

Context

1,048,576 tokens

Max output

1,048,576 tokens

Streaming

Supported

Tools

Supported

Vision

Supported

Supported API forms

/v1/chat/completions · /v1/messages

Send a request

/v1/chat/completions

curl https://api.wokey.ai/v1/chat/completions \
  -H "Authorization: Bearer $WOKEY_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"kimi-k3","messages":[{"role":"user","content":"Hello"}]}'

/v1/messages

curl https://api.wokey.ai/v1/messages \
  -H "x-api-key: $WOKEY_API_KEY" \
  -H "anthropic-version: 2023-06-01" \
  -H "Content-Type: application/json" \
  -d '{"model":"kimi-k3","max_tokens":1024,"messages":[{"role":"user","content":"Hello"}]}'

Use Kimi K3 in your tools

The gateway is https://api.wokey.ai (OpenAI-compatible clients usually take https://api.wokey.ai/v1), and the model ID is kimi-k3. Open the guide for the client you use.

• Use Kimi K3 in Claude Code: Anthropic Messages · base URL and API key setup
• Use Kimi K3 in Claude Desktop: Anthropic Messages · base URL and API key setup
• Use Kimi K3 in opencode: OpenAI Chat Completions · base URL and API key setup
• Use Kimi K3 in OpenClaw: OpenAI Chat Completions · base URL and API key setup
• Use Kimi K3 in Hermes Agent: OpenAI Chat Completions · base URL and API key setup
• OpenRouter alternative: Wokey vs OpenRouter: Compare token prices, payment fees, API forms, and upstream sources.
• Verifiable AI API: which models and API forms carry response proofs: Response proofs currently cover only raw passthrough Claude Messages and GPT Responses on official routes.

Example production call

These usage values come from a completed Kimi K3 call that succeeded on its first attempt and was reconciled with its bill.

Total input

76023

Output

574

Cached share of input

99.3%

Call duration

25.60 s

Chat Completions usage

Compiled from the billing record into API field examples. prompt_tokens is total input, including cached tokens.

{
  "model": "kimi-k3",
  "usage": {
    "prompt_tokens": 76023,
    "prompt_tokens_details": {
      "cached_tokens": 75520
    },
    "completion_tokens": 574,
    "total_tokens": 76597
  }
}

Token accounting

These are the final per-million-token rates recorded for this call.

Billed amount

$0.009833

Cost = sum of tokens × their rate ÷ 1,000,000, rounded to six decimals after summing.

Source of this call

This call used the Volcengine channel at ark.cn-beijing.volces.com with API Key authorization. Usage was metered upstream and the request returned HTTP 200.

Verify Prompt Cache

1. Keep the shared prefix of long prompts identical and reuse context according to the selected API’s cache rules.
2. Check prompt_tokens_details.cached_tokens; zero means no cache reads were recorded for this call.
3. Uncached input = prompt_tokens minus cached tokens. Price each bucket separately and add the output charge.

Frequently asked questions

Related models and documentation

• Kimi K2.8 Preview
• API documentation
• All models and pricing
• Provider Node