Short answer
Moonshot AI released Kimi K3.1 Pro on November 4, 2026. The Highspeed variant was announced on November 10, alongside the model's arrival on Amazon Bedrock. The model ID is kimi-k3-1-pro-highspeed.
The numbers, all from the launch materials:
- Speed: Highspeed generates output two to three times faster than K3.1 Pro.
- Context: 1 million tokens, the same as K3.1 Pro.
- Bedrock standard tier (per 1 million tokens, Global cross-Region profile): $4.00 input, $20.00 output, $0.20 cache read, $5.00 cache write (30 minutes).
- Reasoning levels:
none,low,medium,high,xhigh,max, defaulting tomax. - Capabilities: tool calling, structured output, image input, streaming, configurable reasoning.
- APIs: OpenAI-compatible Chat Completions and Responses, or Bedrock's Converse and Invoke.
kimi-k3-1-pro-highspeed is not on TeamoRouter yet. When it goes live, the key and base URL you already have will work — just change model to kimi-k3-1-pro-highspeed. We also don't carry any other Moonshot model, so there is no same-vendor substitute to switch to in the meantime.
What was announced
K3.1 Pro is a post-trained version of Kimi K3: the architecture and context window are unchanged — still a 3T-class open-weight model — but it was trained with substantially more post-training compute. Moonshot says that extra compute went into long-horizon agentic work and cutting factual errors, with improvements showing up in financial modeling, office documents, and code for game engines, 3D scenes and CAD scripts.
Highspeed is not a different model. It is a faster version of the same one: two to three times the output rate, with the same context window, capabilities and APIs. It is available on Amazon Bedrock from November 10 through three inference profiles:
- Global cross-Region: routes to any supported commercial AWS Region.
- US cross-Region: keeps routing inside the US.
- India cross-Region: keeps routing inside India, splitting requests between Mumbai and Hyderabad.
The last two exist for data-residency requirements. Separately, inference data stays inside the AWS data boundary, is not shared with Moonshot AI and is not used to train the model. Zero data retention is always on for inference requests, and AWS operators cannot view prompts or completions.
If you already run Kimi K3 on Bedrock, switching is a model-ID change. AWS recommends Chat Completions over the other options.
What it supports, and what it doesn't
What works, per the official material:
- Tool calling: same interface as K3, no separate endpoint.
- Structured output: constrain responses to a schema such as JSON.
- Configurable reasoning: six levels, covered below.
- Image input: you can send images.
- Streaming: tokens come back as they are generated.
The gaps are worth knowing before you build on it:
- No video input. Bedrock doesn't support video for this model.
- Converse rejects PDF and HTML attachments. Handle those documents some other way.
- Converse has a multi-turn trap: if earlier turns' reasoning content is included in the request, it can fail with an internal server error. Frameworks like LangChain and Strands Agents hit this in their default configurations, so strip prior reasoning blocks before multi-turn Converse calls.
- Explicit prompt caching doesn't work on Converse — only on Responses, Chat Completions and Invoke. Each cache breakpoint needs at least 256 tokens of stable content ahead of it.
- Output tokens eat quota: on
bedrock-runtime, every output token consumes 10 tokens from your account quota. That's a capacity constraint, separate from the bill.
Pricing
First the scope: these are Amazon Bedrock standard-tier rates, split by inference profile, per 1 million tokens in USD. Pricing for Moonshot's own API is not in these launch materials, so we're not guessing at it here.
| Inference profile | Input | Output | Cache read | Cache write (30 min) |
|---|---|---|---|---|
| Global cross-Region | $4.00 | $20.00 | $0.20 | $5.00 |
| US cross-Region | $4.40 | $22.00 | $0.22 | $5.50 |
| India cross-Region | $4.40 | $22.00 | $0.22 | $5.50 |
Two other tiers exist and are selected per request:
- Flex: half the standard rate, in exchange for potentially slower responses. A good fit for evaluations, summarization and background agent work.
- Priority: a 75% premium for tighter latency on latency-critical calls.
Two caching details matter for cost: cache reads are discounted and don't count toward the input tokens-per-minute quota, while cache writes cost more than ordinary input. Point caching at context that genuinely repeats — system instructions, tool definitions, repository conventions — not at one-off content.
One more cost trap: the default reasoning level is max. Run with no settings at all and every request is billed at the deepest setting.
Choosing a reasoning level
The supported values are none, low, medium, high, xhigh and max, with max as the default.
| Level | Behaviour |
|---|---|
none |
Reasoning off for the request |
low |
Least deliberation, lowest latency and token use |
medium |
Moderate |
high |
Deep |
xhigh |
Deeper |
max (default) |
Deepest, highest token consumption |
The trade-off at lower levels is explicit: less latency and fewer tokens, but weaker task performance. For cost-sensitive workloads the first move is to drop off the default max and find, through your own evaluations, the lowest level that still meets your quality bar. The reverse also holds: for long-horizon agent work and complex multi-step reasoning, don't force it down to none to save money.
How to call it
The model isn't on TeamoRouter yet, so the snippet below shows how it will be called once it is live. You can write the integration against it now, but it won't run yet — don't treat it as available when you test.
OpenAI-compatible format, with your existing base URL and key:
from openai import OpenAI
client = OpenAI(base_url="https://api.teamorouter.com/v1", api_key="YOUR_KEY")
resp = client.chat.completions.create(
model="kimi-k3-1-pro-highspeed",
messages=[{"role": "user", "content": "Make this CAD script parametric: ..."}],
)
print(resp.choices[0].message.content)
The same OpenAI-compatible interface also offers the Responses form:
curl https://api.teamorouter.com/v1/responses \
-H "Authorization: Bearer $TEAMOROUTER_KEY" \
-H "Content-Type: application/json" \
-d '{"model": "kimi-k3-1-pro-highspeed", "input": "Explain long-horizon agentic work in one sentence."}'
On the vendor-native side — AWS's own entry point — you use Bedrock's Converse API:
import boto3
client = boto3.client("bedrock-runtime", region_name="us-east-1")
resp = client.converse(
modelId="kimi-k3-1-pro-highspeed", # use the full modelId shown in the Bedrock console
messages=[{"role": "user", "content": [{"text": "Make this CAD script parametric: ..."}]}],
inferenceConfig={"maxTokens": 2048},
)
print(resp["output"]["message"]["content"][0]["text"])
Both Converse caveats from above apply here: strip prior reasoning blocks on multi-turn requests, and it won't accept PDF or HTML attachments. Reasoning-level controls are also fully available only on Responses, Chat Completions and Invoke.
What to use in the meantime
kimi-k3-1-pro-highspeed is not on TeamoRouter yet. Calling that ID today will fail, and it won't be your key or base URL at fault.
Once it's live, the same key and the same base URL are all you need — change model to kimi-k3-1-pro-highspeed and nothing else in your code has to move. It runs on the standard OpenAI-compatible path, not a separate endpoint.
Since we carry no other Moonshot model either — no K3, no K3.1 Pro — there's no same-vendor stand-in worth hunting for. Keep running whatever you already have in production, and switch when kimi-k3-1-pro-highspeed appears in the catalog. The switch itself is the same one-line change.
How to pay
No subscription: top up your balance in the console from $5, and each request is billed by the tokens it uses. Cards are supported, as are USDT, USDC and ETH. The same key also calls Claude and GPT models.
New accounts start with a $0 balance — top up before your first request. Top up from $5. If a request returns 402 (400 on the OpenAI format) with "wallet balance is insufficient", the balance is empty, not your base URL or key. Top up and retry.
FAQ
Is Kimi K3.1 Pro Highspeed out?
Yes. K3.1 Pro launched on November 4, 2026, and the Highspeed variant was announced on November 10 with its Amazon Bedrock availability.
What's the model ID?
kimi-k3-1-pro-highspeed. On Bedrock, pass the full modelId shown in the console; on the OpenAI-compatible endpoint, use the ID above.
How much does it cost?
On Bedrock's standard tier, per million tokens: $4.00 input, $20.00 output, $0.20 cache read and $5.00 cache write on the Global profile; $4.40 / $22.00 / $0.22 / $5.50 on the US and India profiles. Flex is half the standard rate; Priority adds a 75% premium.
Is it a different model from K3.1 Pro?
No. It's the same model at a higher output rate — two to three times faster — with the same context window, capabilities and APIs.
How large is the context window?
1 million tokens. K3.1 Pro is a post-trained Kimi K3, and it keeps K3's architecture and context window.
Does it take video?
No. Video input isn't supported, and Converse rejects PDF and HTML attachments.
Can I call it on TeamoRouter now?
Not yet. kimi-k3-1-pro-highspeed isn't in our catalog, so calls to it will fail today. When it goes live, the same key and base URL work — just change model. We don't carry any other Moonshot model to use in the interim.
Related reading
- Kimi K3.1 Pro Highspeed: access and setup: what you need before you switch
- Kimi K3.1 Pro Highspeed compared with the alternatives: scenario-by-scenario, with pricing