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

Kimi K3.1 Pro API Benchmarks: Providers, Speed & Cost

Published on 2026.11.17 by DeepInfra
Kimi K3.1 Pro API Benchmarks: Providers, Speed & Cost

Kimi K3.1 Pro is Moonshot AI’s new open-weights flagship, and it tops the Artificial Analysis Intelligence Index among open-weights models of its size. It is also a large, talkative reasoning model, so where you run it changes both the bill and the wait. This guide compares the Artificial Analysis data for the 16 endpoints from 14 providers covered here, price first.

Kimi K3.1 Pro Provider Comparison Data

Prices are per 1M tokens. Blended price uses Artificial Analysis’s 7:2:1 weighting of cache-hit, input and output tokens.

Provider (endpoint)InputOutputBlended (7:2:1)Error rateOutput speed (t/s)Time to first answer token (s)Cost per Index taskBest for
DeepInfra$1.90$9.50$1.460.0%46.9109.82$4.39Cost with reliability: lowest price of any endpoint with a 0% error rate; Flex tier
PrimaLabs$1.80$9.00$1.3915.4%109.821.16$2.94Lowest list price
Modal$2.00$10.00$1.540.0%118.044.64$3.19Cache-heavy agent runs
Baseten$2.00$10.00$1.540.0%183.025.53$3.26Throughput at list price
Databricks$2.00$10.00$1.541.9%111.769.25$3.29Databricks-native stacks
DigitalOcean$2.00$10.00$1.540.0%200.825.35$3.42Speed and latency at list price
Wafer$2.00$10.00$1.540.0%112.381.11$3.50Low time to first token (1.41 s)
Kimi (first-party)$2.00$10.00$1.548.0%31.0152.69$3.68Baseline; Moonshot’s own API
Parasail$2.00$10.00$1.540.0%112.154.07$4.22Mid-pack all-rounder
Nebius$2.00$10.00$1.542.0%94.661.52$4.38Mid-pack all-rounder
SiliconFlow$2.00$10.00$1.540.0%45.5105.62$4.68List-price fallback
Inco$3.00$15.00$2.315.8%134.238.91$4.79Faster than list price, below fast tiers
Amazon Bedrock (High Speed)$4.00$20.00$2.940.0%218.224.93$4.37AWS-native fast tier
Kimi High Speed$4.00$20.00$2.944.9%69.474.05$4.42Moonshot’s own fast tier
LithosAI (ULTRA)$6.00$30.00$4.410.0%258.517.87$6.32Fastest output and first answer
Inco (FAST)$6.00$30.00$4.620.0%187.027.22$9.55Premium speed tier

Methodology. Data from the Artificial Analysis providers page for Kimi K3.1 Pro at max reasoning effort, pulled November 17, 2026. Artificial Analysis uses its “Performance v3” method: headline speed and latency average single-request and multi-turn agentic workloads, and error rate pools those runs. Blended price weights cache-hit, input and output tokens 7:2:1. Cost and time per task come from Artificial Analysis’s Intelligence Index workload and depend on each endpoint’s measured cache hit rate as well as its prices. DeepInfra’s figures are its current promotional price (5% off its $2.00 / $10.00 list price). Figures change over time. Pricing as of 2026-11-17.

Kimi K3.1 Pro API Review Summary

  • Lab and release: Moonshot AI, released October 23, 2026; open weights followed in early November.
  • Intelligence: 46 on the Artificial Analysis Intelligence Index v5.0.1, #1 of 29 open-weights models of similar size (class median 23).
  • Price on DeepInfra: $1.90 input / $9.50 output / $0.19 cached per 1M tokens (5% off list); Flex $1.52 / $7.60 / $0.152.
  • Blended price: $1.46 per 1M on DeepInfra, the lowest of any benchmarked endpoint with a 0% error rate.
  • Reliability: DeepInfra recorded a 0.0% error rate over Artificial Analysis’s measurement window.
  • Speed leaders at list price: DigitalOcean (200.8 t/s, 25.35 s to first answer token) and Baseten (183.0 t/s, 25.53 s).
  • Verbosity: 113M output tokens to run the Intelligence Index, against a class median of 104M.
  • Context and output: 1,048,576-token context; up to 943,718 output tokens on DeepInfra (65,536 by default if max_tokens is unset).
  • Modality: text and image in, text out.
  • Size: Mixture-of-Experts, 2.8T total and 104B active parameters (Hugging Face).
  • License: custom Kimi K3.1-Pro License; commercial use allowed, with extra terms for very large model-as-a-service operators and very large consumer products.
  • Providers covered: 16 endpoints from 14 providers.

About Kimi K3.1 Pro

Kimi K3.1 Pro is post-trained from Kimi K3 with much larger reinforcement-learning compute. The main architecture change is in attention: K3’s full-attention Gated MLA layers are replaced with Gated DeepSeek Sparse Attention for cheaper long-context work, alongside its KDA layers. It keeps the 1M-token context and reads images through a 401M-parameter vision encoder.

Moonshot’s own results (vendor-reported, run at max reasoning effort) put it at the top of its comparison table on agentic tasks such as HLE-Full with tools (70.0), BrowseComp (92.1) and AutomationBench-AA (73.3). Closed models still lead on reasoning without tools (HLE-Full 51.1 vs Claude Opus 5.5’s 59.8), on Terminal-Bench 4.0 by a hair (65.7 vs 66.4) and on the cyber-security rows. Against its predecessor, Moonshot reports Kimi Code Bench 3.0 Long-horizon at 61.2 vs Kimi K3’s 55.95. Our Kimi K3 API provider analysis covers the previous model.

Two serving details shape provider choice. First, the model reasons at length: Artificial Analysis counts 113M output tokens to complete its index, and reasoning tokens bill as output, so the output price dominates cost. Second, latency to the first answer is long on every endpoint (17.87 s to 152.69 s), so time to first answer token matters far more than time to first token.

DeepInfra: The Overall Recommended Provider

DeepInfra is the recommended Kimi K3.1 Pro API on price and reliability. At $1.90 input and $9.50 output per 1M tokens it is 5% below the $2.00 / $10.00 list price that 9 of the 16 benchmarked endpoints charge, and its $1.46 blended price is the lowest of any endpoint that recorded a 0% error rate. Among the providers Artificial Analysis benchmarks, only PrimaLabs lists a lower price.

The Flex service tier takes the price lower: $1.52 input, $7.60 output and $0.152 cached. That output rate is 24% below list. At the same 7:2:1 weighting Artificial Analysis uses, Flex works out to about $1.17 blended (our calculation). Send "service_tier": "flex"; a Flex request waits up to 10 minutes for capacity, then returns HTTP 429, and a rejected request isn’t charged. A Priority tier is also available at 1.5× the promotional rate.

Because output dominates this model’s bill, the per-token gap adds up. A batch job that generates 100M output tokens costs $950 on DeepInfra Standard, $760 on Flex, and $1,000 at list price.

If response speed is your product, test DigitalOcean or Baseten at the standard list price; for batch, agents and cost-sensitive work, DeepInfra’s per-token gap compounds.

  • Blended Price: $1.46 per 1M ($1.90 input, $9.50 output, $0.19 cached); Flex about $1.17
  • Output Speed: 46.9 t/s
  • Latency (TTFT): 2.09 s; 109.82 s to first answer token
  • Optimal Workload: batch reasoning, background agents and evaluation runs where price and a clean error rate matter more than seconds to first answer

The Kimi K3.1 Pro model page lists JSON mode, structured output, tool calling, image input, zero data retention and private endpoints. The API is OpenAI-compatible at https://api.deepinfra.com/v1/openai/chat/completions with Authorization: Bearer $DEEPINFRA_API_KEY; parameters are on the API page.

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": "Plan the migration steps for this schema change."}],
    temperature=1.0,
    top_p=0.95,
    max_tokens=32768,
    extra_body={
        "reasoning_effort": "high",
        "service_tier": "flex",  # remove for standard tier
    },
)
print(response.choices[0].message.content)
copy

DigitalOcean: Speed and Latency Leader at List Price

DigitalOcean is the fastest endpoint at the standard $2.00 / $10.00 price: 200.8 t/s output and 25.35 seconds to the first answer token, with a 0% error rate. It finishes an Intelligence Index task in 6.0 minutes and costs $3.42 per task.

  • Blended Price: $1.54 per 1M
  • Output Speed: 200.8 t/s
  • Latency (TTFT): 1.71 s; 25.35 s to first answer token
  • Optimal Workload: interactive coding assistants and chat where users wait for the answer

Baseten: High-Throughput Alternative

Baseten is close behind at 183.0 t/s and 25.53 seconds to first answer, also at list price with a 0% error rate. A 99.6% cache hit rate on Artificial Analysis’s agentic workload brings its cost per task to $3.26, the third-lowest in the set.

  • Blended Price: $1.54 per 1M
  • Output Speed: 183.0 t/s
  • Latency (TTFT): 1.44 s; 25.53 s to first answer token
  • Optimal Workload: throughput-heavy agent pipelines that also need low latency

LithosAI ULTRA and Amazon Bedrock High Speed: Premium Speed Tiers

If speed matters more than price, two fast tiers lead the table. LithosAI ULTRA has the highest output speed (258.5 t/s) and the shortest time to first answer (17.87 s) at $6.00 / $30.00, three times list price. Amazon Bedrock High Speed reaches 218.2 t/s and 24.93 s at $4.00 / $20.00. Both recorded 0% error rates.

  • Blended Price: $4.41 (LithosAI ULTRA); $2.94 (Bedrock High Speed)
  • Output Speed: 258.5 / 218.2 t/s
  • Latency (TTFT): 1.75 / 4.00 s; 17.87 / 24.93 s to first answer token
  • Optimal Workload: latency-critical products with budget for a speed premium

PrimaLabs: Lowest Listed Price

PrimaLabs lists the lowest price on Artificial Analysis ($1.80 / $9.00, $1.39 blended) and the lowest cost per Index task ($2.94), with 109.8 t/s and 21.16 seconds to first answer. Its 15.4% error rate over the measurement window is the highest in the set, so plan retries or a fallback if you route to it.

  • Blended Price: $1.39 per 1M
  • Output Speed: 109.8 t/s
  • Latency (TTFT): 0.95 s; 21.16 s to first answer token
  • Optimal Workload: cost-first workloads with retry logic in place

Kimi (Moonshot First-Party API): The Baseline

Moonshot’s own API charges the $2.00 / $10.00 list price and measured 31.0 t/s with 152.69 seconds to first answer token and an 8.0% error rate. Moonshot also sells Kimi K3.1 Pro Highspeed at $4.00 / $20.00 (69.4 t/s on Artificial Analysis). The first-party API is the reference for pricing and for Moonshot-specific features.

  • Blended Price: $1.54 per 1M (Highspeed $2.94)
  • Output Speed: 31.0 t/s (Highspeed 69.4)
  • Latency (TTFT): 4.03 s; 152.69 s to first answer token
  • Optimal Workload: teams that want Moonshot’s own API and products

Technical Deep-Dive: What Developers Need to Know

1. Time to First Answer Token Is the Latency Number That Matters

Time to first token runs from 0.95 to 6.54 seconds across the 16 endpoints, but time to first answer token runs from 17.87 to 152.69 seconds, because the model reasons before answering. For anything user-facing, compare providers on time to first answer, and consider a lower reasoning_effort for interactive calls: DeepInfra’s API accepts none through max, and Artificial Analysis’s figures are all measured at max.

2. Output Price Drives the Bill

At 113M output tokens to run Artificial Analysis’s index (class median 104M), Kimi K3.1 Pro is somewhat verbose, and its output price is five times its input price. On output-heavy work, the output rate is the number to compare: $9.50 on DeepInfra, $7.60 on Flex, $10.00 at list price and $20.00–$30.00 on fast tiers.

3. Cost per Task Depends on Your Cache Hit Rate

Artificial Analysis’s blended price assumes 70% of tokens are cache hits, and its cost per task reflects each endpoint’s measured cache hit rate on a long agentic workload. If your traffic is mostly fresh prompts or output-heavy reasoning, per-token prices predict your bill better than the per-task figure. Use prompt_cache_key on DeepInfra to keep repeated context cached across agent turns.

4. Output Limits and Long Context

DeepInfra serves the full 1,048,576-token context and allows up to 943,718 output tokens. If you don’t set max_tokens, it defaults to the smaller of the remaining context and 65,536.

5. API Feature Parity

All the providers above serve the same open weights through OpenAI-compatible APIs, so you can route cost-sensitive traffic to one endpoint and latency-sensitive traffic to another by changing only the base URL and model name.

Conclusion

  • Lowest price with a 0% error rate: DeepInfra ($1.90 / $9.50, $1.46 blended; Flex $7.60 output)
  • Speed and latency at list price: DigitalOcean (200.8 t/s, 25.35 s to first answer); Baseten close behind
  • Fastest regardless of price: LithosAI ULTRA (258.5 t/s, 17.87 s) and Amazon Bedrock High Speed
  • Lowest listed price: PrimaLabs, with a 15.4% error rate to plan around
  • Baseline: Moonshot’s first-party Kimi API at $2.00 / $10.00

For full pricing scenarios, see the Kimi K3.1 Pro pricing guide; for the model itself, the Kimi K3.1 Pro model overview. Our roundup of the best open-source models for coding puts K3.1 Pro next to its open-weights peers. To start, call Kimi K3.1 Pro on DeepInfra, or deploy it on a private endpoint for dedicated GPU capacity.

Related articles
Best OpenClaw Alternatives: Hermes Agent, ZeroClaw & NemoClawBest OpenClaw Alternatives: Hermes Agent, ZeroClaw & NemoClaw<p>OpenClaw has 362,000 GitHub stars and a skill marketplace with over 44,000 community contributions. That kind of adoption doesn&#8217;t happen by accident. Still, the same teams running it in production keep running into the same complaint: the model list is fixed. OpenClaw&#8217;s guided setup wizard covers OpenAI, Anthropic, Google, DeepSeek, and local Ollama. You can [&hellip;]</p>
Best MiMo-V2.5 API Providers RankedBest MiMo-V2.5 API Providers Ranked<p>Executive Summary: Selecting the right API provider for Xiaomi&#8217;s MiMo-V2.5 is critical for optimizing production workflows. Based on the benchmark research, DeepInfra is the best provider for raw speed and low latency (130+ tokens/second), while Xiaomi&#8217;s first-party API is the most cost-effective, offering unmatched prompt caching discounts. This guide breaks down the model&#8217;s MoE architecture [&hellip;]</p>
MiMo-V2.6-Pro Guide: Performance, Pricing & ProvidersMiMo-V2.6-Pro Guide: Performance, Pricing & Providers<p>API Review Summary (October 2026) Attribute Value Developer Xiaomi Release Date September 21, 2026 Model Type Reasoning model (open weights) Artificial Analysis Intelligence Index 30 — ranks #3 of 21 models in AA&#8217;s “Large” (&gt;150B parameter) class, where the class median is 23 Speed 53.4 output tokens/sec (class median: 73.9 t/s) Latency (TTFT) 3.16s (class [&hellip;]</p>