Kimi K3 API Pricing: 1M Context Cost Impact
Kimi K3 launches at $3 input, $0.30 cached input, and $15 output per 1M tokens. Here is the pricing impact for coding agents.
By AI Pricing Guru Editorial Team
AI Pricing Guru articles are maintained by the editorial workflow behind the site: daily pricing snapshots, provider source checks, and review passes for model launches, subscription limits, and billing changes.
Moonshot AI has published pricing for Kimi K3, a new flagship Kimi model aimed at long-horizon coding, end-to-end knowledge work, tool use, and large-context agent workflows.
The headline rate card is simple: $3.00 per million input tokens, $0.30 per million cache-hit input tokens, and $15.00 per million output tokens, with a 1,048,576-token context window. Kimi’s docs position K3 as an always-reasoning model with configurable reasoning effort, automatic context caching, tool calls, JSON mode, structured output, partial mode, and dynamic tool loading.
This is a real pricing event because Kimi K3 lands directly in the same buying conversation as Claude Sonnet, Gemini Pro, GLM-5.2, GPT-5.6 Sol, and cheaper open-model routes. It is not the cheapest model in the market. Its argument is a premium long-context coding and agent model at a price below many top closed-model routes.
For live comparison, keep the Together AI pricing page, OpenAI pricing page, Anthropic pricing page, and AI token calculator open. For adjacent long-context economics, compare this with our earlier GLM-5.2 pricing impact analysis.
What Changed
The July 16-17 alerts picked up two Kimi K3 signals: Moonshot’s official Kimi K3 pricing documentation and third-party performance and price analysis from Artificial Analysis. The official docs now list Kimi K3 as a separate flagship pricing tier rather than another K2.x variant.
The confirmed Kimi K3 pricing table is:
| Item | Kimi K3 detail |
|---|---|
| Standard input | $3.00 / 1M tokens |
| Cache-hit input | $0.30 / 1M tokens |
| Output | $15.00 / 1M tokens |
| Context window | 1,048,576 tokens |
| Model ID | kimi-k3 |
| Positioning | Long-horizon coding and end-to-end knowledge work |
Kimi also says K3 supports automatic context caching, tool calls, JSON mode, structured output, partial mode, internet search, tool choice constraints, and dynamically loaded tools. That last feature matters for cost: large tool inventories can inflate prompts quickly, so dynamically loading only the relevant tools is one of the cleaner ways to control context spend.
Pricing Comparison
Here is the practical comparison against other agent and long-context routes in our tracked data.
| Model | Input | Cached input | Output | Pricing read |
|---|---|---|---|---|
| DeepSeek V4 Pro | $0.435 | $0.003625 | $0.87 | Budget coding and reasoning route |
| Together Kimi K2.7 Code | $0.95 | $0.19 | $4.00 | Cheaper Kimi-coded workload route |
| GLM-5.2 | $1.40 | $0.26 | $4.40 | Open-weight long-context coding model |
| Claude Sonnet 5 | $2.00 | $0.20 | $10.00 | Premium Claude production route |
| Kimi K3 | $3.00 | $0.30 | $15.00 | Flagship Kimi long-context route |
| Gemini 3.1 Pro | $2.00 | $0.20 | $12.00 | Google long-context premium route |
| GPT-5.6 Sol | $5.00 | $0.50 | $30.00 | Flagship OpenAI preview route |
Kimi K3 is cheaper than GPT-5.6 Sol and GPT-5.5 on both input and output. It is roughly in the Claude Sonnet and Gemini Pro price band, but with a higher listed output rate than Claude Sonnet 5 and Gemini 3.1 Pro. It is much more expensive than DeepSeek V4 Pro, Kimi K2.7 Code, or GLM-5.2 on raw token price.
That means Kimi K3 should not be treated as the default cheap model. It should be evaluated as an escalation tier for tasks where long context, reasoning, tool use, or coding quality can reduce retries and human cleanup.
Cost Example
Take a coding-agent run with:
- 1M fresh input tokens
- 4M cached input tokens
- 500K output tokens
At current listed rates, the rough model bill is:
| Model | Estimated run cost |
|---|---|
| DeepSeek V4 Pro | $0.88 |
| Together Kimi K2.7 Code | $3.71 |
| GLM-5.2 | $4.64 |
| Claude Sonnet 5 | $7.80 |
| Gemini 3.1 Pro | $8.80 |
| Kimi K3 | $11.70 |
| GPT-5.6 Sol | $22.00 |
This is the economic shape buyers should notice. Kimi K3 is not trying to beat DeepSeek or GLM-5.2 on pure price. It is trying to offer a stronger flagship route that still comes in far below GPT-5.6 Sol for long agent traces.
The cache-hit price is especially important. A 1M-token context window can become expensive if every turn is treated as fresh input. If stable system prompts, repository summaries, tool schemas, and long reference material can hit cache, Kimi K3’s effective price changes dramatically.
What This Means
For coding-agent builders, Kimi K3 belongs in the same eval suite as Claude Sonnet, GPT-5.6 Sol, Gemini Pro, GLM-5.2, and DeepSeek. The key metric is not price per token. It is cost per accepted patch, migration, research answer, or tool-completed task.
For teams that already like Kimi routes, K3 creates a clean tiering strategy. Keep Kimi K2.7 Code or another cheaper model for routine code edits, issue triage, search, and summarization. Escalate to K3 for long-horizon work where the full context window and stronger reasoning have a chance to change the result.
For buyers comparing against Claude, the rate is close enough that quality will decide. Claude Sonnet 5 is cheaper in our tracked data, especially on output, but Kimi K3 offers its own 1M-context, tool-calling, and dynamic-tooling story. The winner depends on the repo, task shape, prompt cache behavior, and accepted-output rate.
For buyers comparing against OpenAI, Kimi K3 is meaningfully cheaper than GPT-5.6 Sol on the same synthetic workload shape. That does not make it better, but it does make it a model worth testing before sending every expensive long-context task to a top OpenAI route.
Who Benefits
Agent teams with large tool inventories benefit if Kimi’s dynamic tool loading works well in practice. Tool schemas are a hidden cost center in agent prompts. Loading fewer tools per step can cut tokens, improve tool selection, and reduce failure loops.
Developers building repository-scale coding agents also benefit. A 1M-token window lets the model see more code, logs, design notes, and prior attempts before summarization kicks in. That can be valuable for multi-file refactors and long bug hunts.
Kimi also benefits from clearer premium positioning. K2.7 Code can remain a lower-cost route, while K3 becomes the flagship model for buyers who want stronger long-context reasoning without jumping straight to the highest OpenAI or Claude spend.
Who Should Wait
Simple workloads should wait. Extraction, short support replies, classification, product-copy variants, and small summarization tasks will usually be cheaper on DeepSeek, smaller Gemini routes, Groq-hosted open models, or Kimi K2.7 Code.
Teams without prompt-cache instrumentation should also wait before moving high-volume jobs. At K3’s price, cache hit rate can make or break the bill. Track fresh input, cached input, output, retries, tool-call failures, and accepted-task rate before committing.
Enterprise buyers should verify operational details before depending on K3 for production workflows: rate limits, data handling, regional requirements, support, invoice controls, and uptime. A strong rate card is only part of procurement.
Practical Advice
Benchmark Kimi K3 on tasks that actually use its strengths. Good tests include repo-wide debugging, large-document synthesis, policy-heavy support automation, multi-tool agents, and long-running coding workflows.
Do not compare it only against cheaper Kimi models. Compare it against your current premium fallback. If K3 replaces a GPT-5.6 Sol or top-Claude route often enough, the savings can be meaningful. If it replaces GLM-5.2 or DeepSeek, the quality uplift needs to be obvious.
Design prompts for cache hits. Keep stable instructions, tool schemas, repo summaries, and policy blocks consistent across turns where possible. Kimi K3’s cache-hit input price is one-tenth of its standard input price, so cache discipline is part of the pricing strategy.
Use the AI token calculator before broad rollout. A model with a 1M-token window can make it tempting to pass everything. That is rarely the cheapest answer. Route small tasks to cheaper models and reserve K3 for the context-heavy steps that justify the premium.
Bottom Line
Kimi K3 is a serious long-context pricing event. At $3 input, $0.30 cached input, and $15 output per million tokens, it is not a budget model. It is a premium Kimi route priced below the highest OpenAI tier and close enough to Claude and Gemini premium routes that real evals will matter more than headline rates.
For AI buyers, the right move is to test Kimi K3 on long-context coding and tool-use workloads immediately, but with strict cost tracking. If it reduces retries or handles larger tasks without losing context, it can earn the higher rate. If the workload is short or routine, cheaper models still win.
Sources: Kimi K3 official pricing docs, Kimi K3 quickstart, Artificial Analysis Kimi K3 page, and AI Pricing Guru’s live pricing dataset.