Kimi K3 Price Guide: $3/$15 per 1M Tokens With 1M-Context Flat Pricing

Kimi K3 costs $3.00 per million input tokens and $15.00 per million output tokens on Moonshot's official API, with a discounted $0.30 per million rate when a cache hit is served. Those rates are flat across the model's full 1,048,576-token context window — no length tiering — and apply to the kimi-k3 model ID at https://api.moonshot.ai/v1 as of July 16, 2026, the day Moonshot began rolling K3 out on Kimi web, the Kimi app, and Kimi Code. Consumer chat access inside the Kimi app is included with a free account, and prepaid credit plans start at ¥199 with launch bonuses running through August 11, 2026.

Key Takeaways

• API rate card: $3.00/M input (cache miss), $0.30/M input (cache hit), $15.00/M output, per Moonshot's official K3 pricing docs.
• Flat across 1M context: unlike some long-context tiers on rival APIs, K3 does not charge a premium for prompts above 200K tokens.
• Cache economics: the 10× cheaper cache-hit tier rewards stable system prompts and long document prefixes — the standard agent-loop pattern.
• Consumer app: free to try Kimi K3 Max inside the Kimi iOS app and kimi.com under standard rate limits.
• Subscription: prepaid credit plans start at ¥199 (~$28); launch top-up bonuses run 10%–30% between July 15 and August 11, 2026.
• Open-weight intent: Moonshot has signalled a weights release "in the coming days," but nothing had landed on Hugging Face at publish time. Self-hosting a 2.8T-parameter model is not a serious cost lever for most teams.

Kimi K3 Pricing Breakdown

The primary rate card comes from Moonshot's own pricing page and the launch-night Hacker News thread that mirrored it. All figures are per one million tokens and exclude local sales tax.

Item	Rate (USD per 1M tokens)	Notes
Input, cache miss	$3.00	Standard prompt tokens on a cold request
Input, cache hit	$0.30	Automatic prefix caching on stable prompts
Output	$15.00	Includes reasoning + final answer tokens
Context window	1,048,576 tokens	Flat pricing across the full window
Model ID	kimi-k3	Only supported reasoning_effort is max today
Default max output	131,072 tokens	Configurable up to the full 1M via max_completion_tokens

Sources: Moonshot's K3 pricing page and the K3 quickstart docs, both live as of July 16, 2026. The Hacker News launch thread independently transcribes the same $3/$15/$0.30 figures.

Subscription and top-up pricing is documented separately on Moonshot's Kimi Open Platform and mirrored in launch coverage from tier-A observers:

Tier	Base amount	Launch bonus	Effective credit
Trial	¥99–¥499	+10%	¥108.90–¥548.90
Standard	¥500–¥1,999	+20%	¥600–¥2,398.80
Pro	¥2,000–¥4,999	+25%	¥2,500–¥6,248.75
Team	¥5,000+	+30%	¥6,500+ on a ¥5,000 base
Kimi Code entry plan	¥199 (~$28)	Bonus per top-up tier above	Confirmed by community teardown

The promotion window runs July 15, 2026 through August 11, 2026 per Chubby's launch-day summary. Rates and bonuses may change after the window closes; check the current page before topping up.

What That Costs in Practice

The published rates only become useful once you translate them into workload-shaped numbers. All figures below use Moonshot's stated $3/M input, $0.30/M cache-hit input, and $15/M output; every calculation is spelled out so you can re-run it for your own prompts.

Scenario	Input tokens	Output tokens	Cache state	Cost calculation	Total
Short chat turn	2,000	500	Miss	(2,000 ÷ 1M × $3) + (500 ÷ 1M × $15)	$0.014
Same turn, warm agent loop	2,000	500	Hit	(2,000 ÷ 1M × $0.30) + (500 ÷ 1M × $15)	$0.0081
Long-document Q&A	200,000	2,000	Miss	(200,000 ÷ 1M × $3) + (2,000 ÷ 1M × $15)	$0.63
Same doc, cached prefix	200,000	2,000	Hit	(200,000 ÷ 1M × $0.30) + (2,000 ÷ 1M × $15)	$0.09
Full 1M-context extraction	1,048,576	5,000	Miss	(1,048,576 ÷ 1M × $3) + (5,000 ÷ 1M × $15)	$3.22
Coding agent, heavy reasoning	20,000	40,000	Hit on prefix	(20,000 ÷ 1M × $0.30) + (40,000 ÷ 1M × $15)	$0.606

Two caveats matter. First, K3 always runs in thinking mode with reasoning_effort=max, so output token counts trend high — the Hacker News launch thread flagged that "obese thinking traces" have historically eroded Chinese open-model price wins. Independent tester LinearUncle reported a single K3 task taking about an hour to finish, versus roughly 30 minutes on Fable 5, which hits your output bill directly. Second, the cache-hit price only applies to prefixes Moonshot's system recognises as stable across requests; live tool responses will not qualify.

How Kimi K3 Pricing Compares

Frontier chat pricing has clustered around a few reference points. K3 lands squarely in the Sonnet-tier band on published rate cards, cheaper than closed frontier models but well above the cheapest open-weight Chinese competitors.

Model	Input ($/M)	Output ($/M)	Cached input ($/M)	Context	Source
Kimi K3	$3.00	$15.00	$0.30	1M	Moonshot docs
GPT-5.6 Sol	~$30.00 output tier	~$30.00	Not yet confirmed	Not yet confirmed	Referenced in HN launch thread
GPT-5.6 Terra	~$2.50	~$15.00	Not yet confirmed	Not yet confirmed	Referenced in HN launch thread
Claude Sonnet 5	On discount (exact rate not in bundle)	Not yet confirmed	Not yet confirmed	Not yet confirmed	Referenced in HN launch thread
GLM-5.2	Roughly 1/3 of K3, per community	Not yet confirmed	Not yet confirmed	1M	Community teardown in HN launch thread
Kimi K2.6 (previous gen)	$0.60	$2.50	Not yet confirmed	256K	Prior-generation writeup on external Medium blog (not linked)

Two takeaways worth quoting. Moonshot has moved K3 into a materially more expensive band than K2.6 — roughly 5× the input rate and 6× the output rate of the prior open-weight generation — while pitching K3 against Opus-tier intelligence rather than the mid-tier. And per-token comparisons flatter K3 more than end-to-end cost comparisons will, because reasoning verbosity determines what a "typical task" actually bills. For a deeper look at how K3 stacks up against Anthropic's flagship on capability, see our earlier analysis of Kimi K3 vs Claude.

Source: @kimmonismus

Free Tier, Limits, and Access

Consumer chat is free. Kimi K3 Max is live on kimi.com, the Kimi iOS app, and the Kimi Code CLI, according to launch-morning confirmations from Mark Kretschmann and Andrew Curran. Basic chat and agent tasks work under a free Kimi account subject to standard rate limits — Moonshot has not published a hard message cap for the free tier as of publish time.

Two model variants are exposed at launch. Per LuminaXspace's rollout note, K3 Max handles chat and single-agent tasks, while K3 Swarm Max supports large-scale parallel processing for multi-agent workloads. Pricing for Swarm Max was not broken out separately on the K3 pricing page at publish time.

API rate limits. The default max_completion_tokens for kimi-k3 is 131,072, configurable up to the full 1,048,576-token window. Vision inputs are supported natively via base64 or ms://<file-id> file references — public image URLs are not accepted. The web search tool is flagged "being updated" and is not recommended for production use.

Third-party access. K3 is available in Battle Mode and Agent Mode on Arena.ai for evaluation. Direct API access remains through Moonshot's own platform; treat other resellers as out of scope for this pricing reference.

Model weights. Multiple observers describe K3 as open-weight and expect a weights release "in the coming days," but as of July 16, 2026 the model had not appeared on Hugging Face. Even once released, Max Weinbach's stress test suggests a 1.5TB-memory Mac Studio cluster is at the edge of what a 2.5T-parameter, 1M-context inference stack can carry — self-hosting is unlikely to be a serious cost lever for most teams.

What We Don't Know Yet

Moonshot published launch-day rates but left several pricing dimensions open:

• Swarm Max pricing. K3 Swarm Max is live in the app for parallel agent fleets, but a per-request or per-agent rate is not yet published.
• Batch API discounts. Moonshot's Batch API exists for prior K2 models; K3-specific batch pricing was not on the K3 rate card at publish time.
• Web search tool pricing. The docs mark the tool "being updated"; billing terms are pending.
• Vision input pricing. The K3 rate card is single-line; whether images consume a token multiplier or fall under the standard $3/M input is not yet confirmed.
• Weights license and hosting cost. No open-weight license terms, quantised community builds, or third-party inference pricing exist yet.
• Post-promo rates. Whether the ¥199 entry plan and top-up bonuses persist past August 11, 2026 is undisclosed.

This page will be amended as those pieces land. For broader model context, see What Is Kimi K3?.

Frequently Asked Questions

How much does the Kimi K3 API cost? Kimi K3 costs $3.00 per million input tokens on a cache miss, $0.30 per million tokens on a cache hit, and $15.00 per million output tokens, according to Moonshot's official pricing page at platform.kimi.ai as of July 16, 2026. Pricing is flat across the full 1,048,576-token context window.

Is Kimi K3 free to use? Kimi K3 is free to try inside the Kimi web app and iOS app under standard usage limits, but the API is pay-as-you-go. Moonshot is running a launch top-up promotion from July 15 to August 11, 2026 that returns 10% to 30% in bonus credits on prepaid balances.

Is Kimi K3 cheaper than Claude Opus 4.8 or GPT-5.6? Kimi K3 matches Anthropic's Sonnet-tier pricing at $3/$15 per million tokens, well below Claude Opus 4.8, and roughly half the per-token cost of GPT-5.6 Sol on output. Real cost also depends on reasoning verbosity, which can erase the headline advantage — see the practice-scenario table above for how quickly output tokens accumulate under reasoning_effort=max.

What is the cheapest way to access Kimi K3? The Kimi consumer app is the cheapest entry point at no incremental cost. For API workloads, the cheapest path is caching stable system prompts to hit the $0.30-per-million cache rate, or subscribing to a Kimi top-up plan starting at ¥199. Self-hosting the open-weight release is technically an option once weights land, but a 2.8-trillion-parameter model needs substantial GPU capacity.

Does Kimi K3 have a subscription plan? Yes. Moonshot sells prepaid credit tiers starting at ¥199, with launch bonuses of 10% at ¥99–¥499, 20% at ¥500–¥1,999, 25% at ¥2,000–¥4,999, and 30% at ¥5,000 or more through August 11, 2026. A ¥5,000 top-up during the promo returns ¥6,500 in usable credit.

How much does a 1M-token Kimi K3 call cost? A full 1,048,576-token input on a cold cache costs about $3.15 at $3 per million tokens. With a warm cache at $0.30 per million tokens, the same input costs about $0.31. Output is billed separately at $15 per million tokens, so a 5,000-token response adds another $0.075.

Is Kimi K3 open source, and does that make it free? Moonshot has signaled open-weight intent for Kimi K3, but weights had not been published on Hugging Face as of July 16, 2026. Even once released, self-hosting a 2.8-trillion-parameter model requires substantial GPU capacity — the API remains the practical option for most builders, and the published $3/$15 rate is what the market will actually pay.

What to Watch Next

Three signals will move this page. First, whether Moonshot ships open-weight K3 files under a permissive license — that opens third-party hosting and reshapes effective pricing across the market. Second, independent measurement of K3's average output-token count on standard tasks — the $3/$15 rate card only wins if reasoning traces are not two or three times longer than GPT-5.6 Terra's. Third, post-August 11 pricing: whether the launch bonuses persist, sunset, or get replaced by a permanent tier.

Building similar long-context or agentic chat features? On kie.ai you can try Claude Opus 4.8, GPT-5.6, and Gemini 3.1 Pro.