Kimi K3 API: pricing, base URL and quick setup
Use Kimi K3 on A2Agent with model ID kimi-k3. The catalog lists a 1M context window. Check token rates below and verify the features your app needs.
Specifications and pricing
| Model ID | kimi-k3 |
|---|---|
| Platform | Moonshot |
| Context window | 1M |
| Tagged for | chat, vision |
| Discount | -20% |
| Input / 1M tokens | |
| Output / 1M tokens |
USD per 1M tokens. Final billing depends on your API key group and metered usage.
What Kimi K3 costs on A2Agent
The official price is $3.00 input and $15.00 output per 1M tokens. On A2Agent it is $2.40 and $12.00. A month of 10M input and 2M output tokens costs $60.00 at the official price and $48.00 on A2Agent.
Good to know
- Moonshot now calls Kimi K3 its previous-generation flagship. The newer Kimi K3.1 Pro is not in the A2Agent catalog.
- The reference rates in the table match the Moonshot international price list as of 2026-11-06.
Where Kimi K3 fits
Long-horizon coding and terminal agents
Moonshot lists long-horizon coding first among the scenarios K3 is built for: long-running engineering tasks, large codebases and terminal tools.
Basis: Moonshot K3 announcement; Terminal Bench 2.1 and DeepSWE results below.
Web research agents
Moonshot reports a high BrowseComp score for K3 with context compression enabled. It suits agents that search, read and cross-check many pages before answering.
Basis: BrowseComp result below, with its context-management condition.
Screenshots and image input
The catalog tags Kimi K3 for vision. Pass screenshots or design mocks alongside code when the task depends on what the page looks like.
Basis: Catalog vision tag; Moonshot says K3 takes text and image input.
Where it is a weaker fit
Moonshot says K3 is sensitive to thinking history: keep reasoning content intact when you continue a conversation, and avoid switching to K3 midway from another model. Reasoning is always on, so short replies still spend reasoning tokens.
Basis: Moonshot K3 announcement and API guide.
Vendor-reported benchmarks
These figures come from the vendor, not from A2Agent tests. Harnesses, reasoning effort and sampling differ between vendors, so do not rank models by them. Checked 2026-11-06.
| Benchmark | Score | Vendor test conditions |
|---|---|---|
| Terminal Bench 2.1 | 88.3 | Kimi Code harness, reasoning effort max, temperature 1.0. Pass rate type not stated. |
| DeepSWE | 67.5 | Kimi Code harness, reasoning effort max. Moonshot gives 67.3 with the mini-SWE-agent harness. Pass rate type not stated. |
| BrowseComp | 91.2 | Context compression triggered at 300K tokens. Moonshot gives 90.4 with a 1M window and no context management. |
| GPQA-Diamond | 93.5 | Reasoning effort max, temperature 1.0. Tool use and sample count not stated. |
How Kimi K3 compares
K3 is the top of the Moonshot line here and the only one with a 1M context. K2.5, K2.6 and K2.7 Code all stop at 256K. It is priced to match, at $3.00 per million input tokens and $15.00 output, against $0.95 and $4.00 for K2.6. That is roughly three times the input rate and close to four times the output rate, which makes the output side the one to watch. A chatty agent on K3 can spend more on output tokens than input tokens. It reads images like the rest of the K2.x chat models. For a lower-cost comparison, if your prompts fit inside 256K, K2.6 gives you the same modality for a fraction of the cost.
These comparisons use catalog context limits, tags and reference rates, not independent quality benchmarks. Rates can change. Check current pricing above and test your workload; a higher price or a larger context window does not establish better results.
| Model | Context | Input / output per 1M | What differs |
|---|---|---|---|
| GLM-5.3 | 1M | — | Both list a 1M window and keep reasoning on with low, high and max effort. Kimi K3 accepts text and image input; GLM-5.3 takes text only. |
| Qwen3.7 MAX | 1M | — | Alibaba calls Qwen3.7-Max proprietary and publishes no weights. Kimi K3 has open weights under the Kimi K3 License, with 2.8T total and 104B active parameters. |
| MiniMax M3 | 1M | — | MiniMax M3 is the smaller open-weight model, with about 428B total and 23B active parameters. Compare the live rates in the table: the two vendors list very different price points. |
Quick setup
curl https://a2agent.me/v1/chat/completions \
-H "content-type: application/json" \
-H "authorization: Bearer sk-your-a2agent-key" \
-d '{
"model": "kimi-k3",
"messages": [{"role": "user", "content": "Reply with OK"}]
}'Using it from a coding client
Worth it when a single prompt has to hold an entire repository or a long document set. Cap the response length in your client, because the $15.00 output rate is where a 1M session gets expensive.
Questions
What is the context window?
1M tokens, four times the 256K of every other Moonshot model sold here.
Why is it so much more expensive?
It lists at $3.00 input and $15.00 output against $0.95 and $4.00 for K2.6. The window is the visible difference; the output rate is where the gap really shows.
Does it read images?
Yes, it carries the vision tag alongside chat.
When should I stay on K2.6 instead?
Whenever the prompt fits in 256K. Same modality, roughly a third of the input rate and a quarter of the output rate.
Billed per token. Top up from $5, no top-up fee.
