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

Kimi K3.1 Pro on DeepInfra: Pricing, Benchmarks & API

Published on 2026.11.17 by DeepInfra
Kimi K3.1 Pro on DeepInfra: Pricing, Benchmarks & API

Moonshot AI’s Kimi K3.1 Pro posts the top scores in its own comparison table on BrowseComp (92.1) and Humanity’s Last Exam with tools (70.0), ahead of the closed frontier models Moonshot tested against, according to the Kimi K3.1 Pro model card (vendor-reported). Independent testing backs up the headline: Artificial Analysis gives it 46 on its Intelligence Index v5.0.1, the highest of 29 open-weights models of its size, against a class median of 23.

Moonshot released K3.1 Pro on October 23, 2026 and published the weights in early November. It’s a 2.8-trillion-parameter Mixture of Experts model with 104B active parameters, a 1M-token context window and native image input. It’s built for long-horizon agentic work: coding agents, browsing and research, and document-heavy office tasks such as finance, legal and spreadsheet work.

Kimi K3.1 Pro is now available on DeepInfra at $1.90 per 1M input tokens and $9.50 per 1M output tokens with the current 5% promotion. That’s below Moonshot’s own API list price of $2.00 / $10.00, and the Flex tier takes output down to $7.60. Artificial Analysis also recorded a 0% error rate on DeepInfra’s endpoint.

BenchmarkKimi K3.1 ProClaude Opus 5.5GPT-6 AstraDeepSeek-V4.1-Flash-1017GLM-5.3
BrowseComp92.190.891.5––
HLE-Full (with tools)70.067.757.258.862.5
AutomationBench-AA73.369.568.566.662.2
SWE Marathon v1.151.746.342.540.042.5
OfficeQA Pro v264.462.254.440.431.1
Terminal-Bench 4.065.766.458.242.341.8
DeepSWE v1.175.574.274.177.066.9
OSWorld V2.179.681.875.869.3–
HLE-Full (no tools)51.159.853.642.643.0*
GPQA-Diamond94.293.996.192.291.7
ExploitBench74.791.0100–54.4

All scores are vendor-reported from the Kimi K3.1 Pro model card. Moonshot ran K3.1 Pro at max reasoning effort and ran the proprietary models “with fallback”; some comparison scores are cited from leaderboards or other labs’ reports. *Text-only subset.

K3.1 Pro is strongest on agentic tasks that use tools. It leads the table on BrowseComp, HLE with tools, AutomationBench, long-running software work (SWE Marathon) and office document QA. In Moonshot’s own internal hallucination test with tools, its error rate fell to 8.3% from Kimi K3’s 23.4%.

It doesn’t win everywhere. Claude Opus 5.5 edges it on Terminal-Bench 4.0 (66.4 vs 65.7) and computer use (OSWorld V2.1), and the closed models are well ahead on reasoning without tools (HLE-Full and HLE-Diamond) and on cybersecurity, where Moonshot says it removed cyber training data from the open-weight release. Among open-weights models, DeepSeek-V4.1-Flash-1017 beats it on DeepSWE v1.1 (77.0 vs 75.5). Moonshot ran every number above, and the model card also includes internal benchmarks that nobody else can reproduce; those are left out of the table here.

What Makes This Model Different

Kimi K3.1 Pro is the first model in the K3.1 series. Moonshot post-trained it from Kimi K3 with “substantially larger post-training compute”, using large-scale joint reinforcement learning across hundreds of domains. That’s where most of the gains over K3 come from.

The architecture is a sparse Mixture of Experts: 2.8T total parameters, 104B active, 896 experts with 16 selected per token plus 2 shared experts, across 93 layers. Each token touches under 4% of the weights, so you get the capacity of a very large model at the per-token compute of a much smaller one.

Sparse attention for long context

The one architectural change from K3 is attention. K3.1 Pro replaces K3’s full-attention Gated MLA layers with Gated DeepSeek Sparse Attention (Gated DSA). The model now mixes 69 KDA layers with 24 Gated DSA layers, which makes long-context modeling cheaper. That matters at a 1M-token context, where long agent traces and large document sets add up fast.

Native vision

K3.1 Pro takes images as well as text. A 401M-parameter MoonViT-V2 vision encoder feeds the language model directly. Moonshot’s chart, document and multimodal results (CharXiv, MMMU-Pro, GDP.pdf) are strongest when the model can also run Python, which fits agent pipelines that read a chart, compute on it and write up the result.

Reasoning control and output length

It’s a reasoning model, and Moonshot ran its evaluations at max effort with temperature 1.0 and top_p 0.95. On DeepInfra you control depth with reasoning_effort (from none, which turns thinking off, up to max). Reasoning tokens bill as output tokens, so set the effort per task: low for simple steps, max for hard agentic work.

If you leave max_tokens unset, DeepInfra uses the smaller of the remaining context and 65,536 tokens. For long reasoning runs, set it explicitly. The ceiling on DeepInfra is 943,718 output tokens.

License

K3.1 Pro ships under the Kimi K3.1-Pro License, a permissive MIT-style license with two conditions. Companies that offer the model as a service with more than $20M in group revenue over 12 months need a separate agreement with Moonshot first. Products with more than 100M monthly active users or more than $20M in monthly revenue must show “Kimi K3.1-Pro” in their interface. Internal use isn’t affected. Read the license if you’re building a large consumer product on it.

Where it’s not the right fit

Artificial Analysis calls K3.1 Pro “somewhat verbose”: it used 113M output tokens to run the Intelligence Index, against a class median of 104M. Deep reasoning at max also takes time on every provider. Faster endpoints exist (DigitalOcean and Baseten lead output speed at the standard price on Artificial Analysis); if response speed is your product, test them. For batch jobs, agents and cost-sensitive work, the lower per-token rate and Flex add up. For simple, high-volume steps, a smaller model is the better tool (see the conclusion).

Getting Started on DeepInfra

Kimi K3.1 Pro is available now in DeepInfra’s text generation catalog. Pricing is per token with no minimums.

TierInputOutputCached input
Standard, promotional (5% off)$1.90$9.50$0.19
Standard, list$2.00$10.00$0.20
Flex (0.8×)$1.52$7.60$0.152
Priority (1.5×)$2.85$14.25$0.285

Per 1M tokens. Pricing as of 2026-11-17. Flex and Priority apply to the promotional rate. Source: model page and pricing page.

The promotion has no published end date, so budget against the $2.00 / $10.00 list price and treat the 5% as upside. Cached input costs 10% of the input rate, which helps agents that resend the same system prompt and tool definitions on every turn.

As an example, an agent workload of 10M input tokens and 2M output tokens costs $19.00 + $19.00 = $38.00 at the promotional rate. On Flex the same job costs $15.20 + $15.20 = $30.40.

Choosing a tier:

  • Flex costs 0.8× standard. Opt in with "service_tier": "flex". A Flex request waits up to 10 minutes for capacity and then returns HTTP 429, and rejected Flex requests aren’t charged. It suits batch jobs, evals and background agents. See how the Flex tier works.
  • Priority costs 1.5× standard. Opt in with "service_tier": "priority" for traffic that shouldn’t queue.

Model specs on DeepInfra:

  • Model ID: moonshotai/Kimi-K3.1-Pro
  • Context window: 1M tokens (1,048,576)
  • Max output: 943,718 tokens; default max_tokens is 65,536 if unset
  • Input: text and images. Output: text
  • reasoning_effort: none, minimal, low, medium, high, xhigh or max
  • JSON mode, structured output and function calling are supported

The endpoint is OpenAI-compatible. Point your client at https://api.deepinfra.com/v1/openai, call /chat/completions and authenticate with Authorization: Bearer $DEEPINFRA_API_KEY. DeepInfra runs a zero-retention data policy and is SOC 2 and ISO 27001 certified. If you need dedicated capacity, you can deploy a private endpoint from the model page.

Python

Get an API key from the DeepInfra dashboard and export it as DEEPINFRA_API_KEY. This example sends a chart image and asks for analysis:

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": [
                {"type": "text", "text": "Summarize the trend in this chart and flag any anomalies."},
                {"type": "image_url", "image_url": {"url": "https://example.com/revenue-chart.png"}},
            ],
        }
    ],
    max_tokens=131072,  # unset defaults to 65,536; the ceiling is 943,718
    temperature=1.0,
    top_p=0.95,         # Moonshot's evaluation settings
    extra_body={
        "reasoning_effort": "high",  # none | minimal | low | medium | high | xhigh | max
        "service_tier": "flex",      # 0.8x price; drop this line for standard
    },
)

print(response.choices[0].message.content)
copy

To try prompts before writing code, open the interactive demo on the Kimi K3.1 Pro model page. The API reference lists every supported parameter, including reasoning_effort, service_tier, response_format and prompt-caching options.

Conclusion

Kimi K3.1 Pro is the strongest open-weights model on Artificial Analysis’s index for its size class, and its vendor results are best on agentic work with tools: browsing and research, long-running coding tasks, and office and document analysis. Teams building research agents, coding agents or document pipelines should evaluate it on their own tasks, since the benchmark table is Moonshot’s. If you’re running Kimi K3 today, K3.1 Pro is the natural next model to test.

Don’t route every step to it. For classification, extraction and other simple steps, DeepSeek-V4.1-Flash on DeepInfra costs a fraction of the price per token. Keep K3.1 Pro for the hard, multi-step work where its results are strongest.

For more context, see our Kimi K3 launch post and Kimi K3 overview, and how K3.1 Pro compares in our roundup of the best open-source models for coding. The rest of this package covers Kimi K3.1 Pro pricing, API benchmarks, the model overview and the best Kimi K3.1 Pro API providers.

Building a startup? DeepStart gives eligible startups about 50 billion free tokens (counted at DeepSeek V4 Flash prices) to spend on any model, including Kimi K3.1 Pro. Or head to the model page, grab an API key and make your first call.

Related articles
Best MiMo-V2.6-Pro API Providers in 2026Best MiMo-V2.6-Pro API Providers in 2026<p>As large language models continue to push the boundaries of context length and multimodal capabilities, Xiaomi&#8217;s MiMo-V2.6-Pro has emerged as a strong entrant among open-weight models. With its native omnimodal architecture — supporting text, image, video, and audio — and a 1,048,576-token context window, it opens up new paradigms for agentic workflows, long-document analysis, and [&hellip;]</p>
Token Verbosity Is the New Pricing WarToken Verbosity Is the New Pricing War<p>With each new model release, we see new benchmark claims and new arguments justifying the value of its tokens. However, when teams compare models for real-world workloads, the sticker price alone hides large per-task differences in the amount of work a model requires to solve a task. Cheaper tokens do not necessarily translate to lower [&hellip;]</p>
GLM-5.3 API Providers: Speed, Latency & CostGLM-5.3 API Providers: Speed, Latency & Cost<p>GLM-5.3 API Review Summary GLM-5.3 is Z AI’s flagship coding and agentic reasoning model, released on August 14, 2026, and available on DeepInfra since launch. The model is built on the same ~753-billion parameter Mixture of Experts (MoE) base architecture as GLM-5.2, with all performance gains derived entirely from scaled post-training rather than architectural changes. [&hellip;]</p>