DeepInfra raises $107M Series B to scale the inference cloud — read the announcement

| Provider | Best For |
|---|---|
| DeepInfra | Lowest price of the providers in this guide ($1.90 / $9.50), Flex and Priority tiers, 0% measured error rate, private endpoints |
| Moonshot AI (Kimi API) | First-party access, plus the Highspeed variant and Moonshot’s own coding and work apps |
| DigitalOcean | Fast responses at Moonshot’s list price, with dedicated GPUs available |
| Baseten | Fast per-token Model API, with dedicated deployments for scale |
| Modal | Per-token API wired into coding agents, with a low cost per agentic task |
| Amazon Bedrock | The Highspeed variant inside an AWS account, with Flex and Priority tiers |
| OpenRouter | One API key across many hosts, with automatic failover |
| Featherless AI | One OpenAI-compatible key for K3.1 Pro and a large open-model catalog |
Quick reference by priority
Kimi K3.1 Pro is Moonshot AI’s open-weight agentic model, released October 23, 2026, with weights published on Hugging Face in early November. It is a mixture-of-experts model with 2.8T total and 104B active parameters, natively multimodal (text and image in, text out) and a 1M-token context window. It was post-trained from Kimi K3 with much larger reinforcement-learning compute, and it replaces K3’s full-attention layers with Gated DeepSeek Sparse Attention for cheaper long-context serving. On Artificial Analysis’s Intelligence Index it scores 46, the highest of the 29 open-weight models of its size. It is also verbose and reasons for a long time before answering, so where you run it shapes both the bill and the wait.
This guide compares the ways to run Kimi K3.1 Pro on what decides the choice: price and pricing tiers, features, deployment model, and who each option suits. Each entry gets one line on measured speed. The full speed and latency comparison is in the Kimi K3.1 Pro API benchmarks, including time to first token versus time to first answer token, which for this reasoning model ranges from about 25 seconds to over two minutes. Speed, blended price and error rates come from Artificial Analysis (reasoning effort max, pulled 2026-11-17), which benchmarks 16 endpoints from 14 providers for this model. Rates and features come from each provider’s own pages. Pricing as of 2026-11-17.
DeepInfra is the best pick when cost per token decides the choice and you want production reliability, service tiers and a path to dedicated capacity.
"service_tier": "flex". A request waits up to 10 minutes for capacity, then returns HTTP 429, and a rejected Flex request is not charged (how Flex works)"service_tier": "priority", for requests that can’t queuemax_tokens is unset, the default is the lesser of the remaining context and 65,536reasoning_effort from none (thinking off) to max, prompt caching with prompt_cache_key and a configurable TTL, up to 4 fallback models per requestKimi K3.1 Pro is verbose: Artificial Analysis counted 113M output tokens to run its Intelligence Index, against a class median of 104M. Output tokens, including reasoning tokens, therefore drive most of the bill, and DeepInfra’s $9.50 output rate ($7.60 on Flex) is where the saving shows. It suits background agents, document and finance pipelines, evaluation runs and any workload that can absorb a longer wait for a lower price. One thing to measure on your own traffic: DeepInfra’s cache hit rate on Artificial Analysis’s agentic workload was 92.6%, below the 98–100% several hosts reached. Cache-heavy agent loops should set prompt_cache_key and check the hit rate before projecting costs.
A first call with the OpenAI SDK:
import os
from openai import OpenAI
client = OpenAI(
api_key=os.environ["DEEPINFRA_API_KEY"],
base_url="https://api.deepinfra.com/v1/openai",
)
response = client.chat.completions.create(
model="moonshotai/Kimi-K3.1-Pro",
messages=[{"role": "user", "content": "Reconcile these two revenue tables and explain each difference."}],
temperature=1.0,
top_p=0.95,
extra_body={"reasoning_effort": "high", "service_tier": "flex"},
)
print(response.choices[0].message.content)
The endpoint is /chat/completions and the auth header is Authorization: Bearer $DEEPINFRA_API_KEY. Full parameters are on the Kimi K3.1 Pro API page.
Moonshot’s own Kimi API is the best pick when you want first-party access, the Highspeed variant, or the model inside Moonshot’s products.
kimi-k3.1-pro and kimi-k3.1-pro-highspeed (Moonshot release post)The first-party API gets new Moonshot features first, and the Highspeed variant is the simplest way to cut response time without leaving Moonshot. It costs more than DeepInfra on every token category, and Artificial Analysis measured an 8.0% error rate on the standard endpoint in its runs, so build in retries.
DigitalOcean Serverless Inference is the best pick when users wait on every answer and you want to stay at Moonshot’s list price.
DigitalOcean is the provider to test for chat UIs and interactive coding agents. It costs about 5% more per token than DeepInfra’s standard rate, and its pricing page lists only a Standard tier for this model.
Baseten’s Model API is the best pick when you want a fast per-token API now and dedicated deployments when volume grows.
Baseten suits product teams that expect to move from shared to dedicated capacity. Its per-token rates match Moonshot’s list price, so it costs more than DeepInfra per token.
Modal’s Model APIs are the best pick when you run K3.1 Pro inside coding agents and want low cost per finished task.
Modal’s high cache hit rate means agent loops that resend long prefixes pay less per task than the list price suggests. Per token it charges Moonshot’s list rate, about 5% above DeepInfra’s.
Amazon Bedrock is the best pick when your workloads already run in AWS and you want the Highspeed variant under AWS governance.
global.moonshotai.kimi-k3.1), 1M contextnone to max (default max)Bedrock keeps data inside the AWS boundary, with regional profiles for US and India data residency. The standard rate is twice Moonshot’s list price for the Highspeed variant. Bedrock’s Flex tier brings it back to $2.00 / $10.00 for workloads that can wait, which is still above DeepInfra’s standard rate.
OpenRouter is the best pick when you want one API key and automatic failover across the hosts that serve K3.1 Pro.
moonshotai/kimi-k3.1-proOpenRouter lists some smaller hosts with lower prices than DeepInfra, served at a range of precisions. Pin a host if output consistency matters. Teams that settle on one host usually go direct for its full feature set, service tiers and support.
Featherless AI is the best pick when you already use its OpenAI-compatible API for many open models and want K3.1 Pro on the same key.
The 256K window is a quarter of the model’s native context, which matters for whole-repository and long-document work. It fits teams consolidating many open models under one provider.
Kimi K3.1 Pro ships under the Kimi K3.1-Pro License, a custom license that allows commercial use with conditions. Model-as-a-service operators with group revenue above $20 million over any 12 months need a separate agreement with Moonshot, and products with over 100 million monthly active users or over $20 million monthly revenue must display “Kimi K3.1-Pro” in their interface. The license names exemptions for internal use and for some access routes. Read the full text, or ask your provider, if either threshold could apply to you.
Speed and latency for this model shift week to week as hosts tune their serving stacks, as covered in why open models get faster after launch. Check the benchmarks guide for current numbers before committing.
For most teams running Kimi K3.1 Pro as an API, DeepInfra is the default choice. It has the lowest price of the providers in this guide at $1.90 input and $9.50 output, and a Flex tier at $7.60 output for anything that can wait. It is the cheapest endpoint Artificial Analysis measures with a 0% error rate, it serves the full 1M context with up to 943K output tokens, and it offers zero data retention and private endpoints, all on an OpenAI-compatible API you can switch to by changing a base URL. For lighter steps in an agent pipeline, DeepSeek V4.1 Flash runs on the same API for a fraction of the price.
For more detail, see the Kimi K3.1 Pro pricing guide for cost scenarios, the API benchmarks for speed and latency, and the model overview for architecture and parameters. If you are moving from Kimi K3, the Kimi K3 pricing and providers guide covers the previous generation. Startups can apply for free credits through DeepStart.
Chat with books using DeepInfra and LlamaIndexAs DeepInfra, we are excited to announce our integration with LlamaIndex.
LlamaIndex is a powerful library that allows you to index and search documents
using various language models and embeddings. In this blog post, we will show
you how to chat with books using DeepInfra and LlamaIndex.
We will ...
GLM-5.3-Flash Pricing, Providers & Cost<p>GLM-5.3-Flash is a model from Z.ai in the GLM-5 family, released on August 26, 2026. It’s a natively multimodal model accepting text and image input (some listings, including OpenRouter, also note video), with text output. Both Artificial Analysis and DeepInfra describe it as a 320B-parameter mixture-of-experts model with 18B active parameters per inference; DeepInfra’s listing […]</p>
Kimi K2.6 API Benchmarks: Latency, TPS & Cost Analysis (2026)<p>About Kimi K2.6 Kimi K2.6 is an open-source frontier model from Moonshot AI, released on April 20, 2026. It is a native multimodal agentic model built for long-horizon coding, autonomous execution, and swarm-based task orchestration. The model uses a Mixture-of-Experts (MoE) architecture with 1 trillion total parameters and 32 billion activated parameters per token, using […]</p>
© 2026 DeepInfra. All rights reserved.