Kimi K3 API pricing: what Moonshot's 2.8T model costs per token, and how it benchmarks

Kimi K3 is Moonshot AI's new flagship language model, released on July 16, 2026. At 2.8 trillion parameters it is the first model in the 3-trillion class with public weights: Moonshot put them on Hugging Face on July 27, 2026, as promised. The launch was loud enough to move stock markets in three countries. Behind the headlines sits a genuinely interesting model: a new attention architecture, a 1M-token context window, native image and video input, and benchmark results that put it next to the best closed models on several agentic tasks. This post covers what Moonshot published, what the numbers say, and where the claims still need independent confirmation.

TL;DR: Kimi K3 launched July 16, 2026 and is live on UnificAlly as moonshotai/kimi-k3 at $2.25 per 1M input tokens and $11.25 per 1M output, and open weights on Hugging Face since July 27, 2026. Headline specs: 2.8T parameters (16 of 896 experts active per pass), 1M-token context, 131k default output, thinking always on. It posts 88.3% on Terminal Bench 2.1 and 91.2% on BrowseComp, and ranks #14 on Arena's English Text board at 1494 Elo as of October 12, 2026, 19 points behind Claude Fable 5 at #3. Moonshot's API now takes reasoning_effort of low, high, or max, with max still the default. Our take: K3 is the open model to beat, and the default to reach for when a run is long and the bill matters.

What Moonshot released

Kimi K3 is a from-scratch architecture, not a K2 fine-tune. The concrete specs from Moonshot's release:

• 2.8 trillion parameters. The first model in the 3-trillion class with an open release. Moonshot has held the open-model scale frontier for nine of the past twelve months, and K3 extends that run.
• Kimi Delta Attention (KDA) plus Attention Residuals. A hybrid linear attention design meant to keep information flowing through long sequences and deep layers. It is what makes the 1M-token window practical rather than decorative.
• Sparser Mixture of Experts. K3 activates 16 of 896 experts per forward pass under Moonshot's Stable LatentMoE framework. Moonshot reports roughly 2.5x the scaling efficiency of Kimi K2: more capability per unit of compute.
• 1M-token context. A single response defaults to 131k output tokens and can be raised to the full 1M. Context caching is automatic; no cache IDs or TTL settings.
• Native vision, including video. Images and video files go into the same request as text. The model uses screenshots as a feedback loop while coding, which shows up in its frontend and game-development results.
• Thinking always on. K3 runs in thinking mode on every request. At launch reasoning_effort accepted only max; Moonshot's API docs and model card now list low, high, and max, and a request that leaves the field out still runs at max.

The full weights went up on Hugging Face as moonshotai/Kimi-K3 on July 27, 2026, the date Moonshot had promised, with the technical report on GitHub. They ship natively quantized, MXFP4 weights with MXFP8 activations, under the Kimi K3 License.

The benchmarks

Moonshot published a comparison table against Claude Fable 5, GPT 5.6 Sol, Claude Opus 4.8, GPT 5.5, and GLM-5.2, all at max reasoning settings. Credit where due: they included benchmarks they lose.

Where K3 wins:

• SWE Marathon: 42.0%, the top score in the table, ahead of Claude Fable 5 at 35.0%. This is the long-run software engineering test, and it is K3's strongest claim: sustained multi-hour coding with minimal supervision.
• Terminal Bench 2.1: 88.3%, just behind GPT 5.6 Sol at 88.8%. The launch post put Fable 5 at 84.6%, but the model card Moonshot published with the weights swaps in Fable 5's best score across harnesses, 88.0%, so that row is a tie.
• BrowseComp: 91.2% and DeepSearchQA: 95.0 f1, both first in the table. K3 is currently the strongest published model at agentic web search.
• OmniDocBench: 91.1% and SpreadsheetBench 2: 34.8%, both first. Document and structured-file work is a quiet strength.

Where it loses:

• HLE-Full: 43.5% against 53.3% for Fable 5. On the hardest frontier knowledge test, the gap is real.
• FrontierSWE: 81.2% against Fable 5's 86.6%.
• GDPval-AA v2: 1668 Elo against 1760 for Fable 5 and 1748 for GPT 5.6 Sol in the launch post, on economically valuable knowledge work. The July 27 model card updates those to 1686, 1747, and 1736: a smaller gap, still a loss.

On Arena's crowd-voted English Text board, K3 sits at #14 with 1494 Elo as of October 12, 2026, on 11,567 votes, and the rating is no longer marked preliminary. It ranked #10 at 1486 in July, so the Elo went up while the rank went down: newer closed models such as Claude Opus 5.5 and Claude Fable 5.1 arrived above it. Claude Fable 5 sits at #3 with 1513, and the only open-weights model ranked higher than K3 is MiMo V2.6 Pro at #12, on just 1,505 votes. K3's confidence interval still spans ranks 6 to 32, so read the number as a neighborhood. For a model anyone can download, it is a remarkable one.

Independent numbers back the picture: Artificial Analysis measures 93.5% on GPQA Diamond and places K3 fourth of 187 models on its Intelligence Index at a score of 57.

What the open release means

The weights landed on July 27 as promised. That makes K3 the first open model at this scale and shrinks the gap between open and closed frontier models to weeks rather than a year. It matters for three groups:

• Teams with data that cannot leave the building. A near-frontier model you can run in your own infrastructure changes what regulated industries can automate.
• Fine-tuners. A 2.8T MoE is expensive to serve, but an open flagship gives labs and enterprises a base no previous open release matched.
• Everyone negotiating API pricing. Open availability puts a ceiling on what closed providers can charge for comparable capability.

The caveat: 2.8 trillion parameters is not a model you run on a workstation. Even with 16-of-896 sparsity, self-hosting K3 is a multi-node project. Moonshot recommends vLLM, SGLang, or TokenSpeed for serving. For most teams the open weights matter less as a deployment path and more as bargaining power and insurance.

Read the license before you build a hosted product on them. The Kimi K3 License allows commercial use, fine-tuning, and redistribution, but a model-as-a-service business with more than 20 million US dollars in revenue over 12 months needs a separate agreement with Moonshot.

How to read the claims

The weights question is settled. Two things still deserve care:

• Most benchmark numbers are Moonshot's own, and its two tables disagree. The model card published with the weights revises several competitor rows from the launch post, as noted above. Independent checks are arriving, and so far they back Moonshot up: the DeepSWE leaderboard, which runs every model on the same mini-swe-agent harness, scored K3 at 69% (plus or minus 5) in its September 22, 2026 update, against Moonshot's own 67.5.
• Max is still the default. Lower effort settings exist now, but a request that does not set reasoning_effort runs full reasoning, and Artificial Analysis measured output speed at 62 tokens per second at launch, below the median for its class. Simple chat and extraction pay latency they do not need unless you turn the effort down.

Kimi K3 API pricing

K3 bills per token on UnificAlly, quoted per million:

Context caching is automatic, so a coding agent that resends a growing transcript every turn pays the $0.225 cache-read rate on most of its input, not the full $2.25. One number to anchor on: an agent step sending 50,000 input tokens and returning 5,000 costs about $0.1125 in plus $0.05625 out. Remember that reasoning tokens bill as output and K3 reasons at max on every request, so real output volume runs higher than the visible answer text. Live rates for every model are on the pricing page.

How to access Kimi K3 today

• UnificAlly. K3 is live as moonshotai/kimi-k3 on the OpenAI-compatible /v1/chat/completions endpoint, so existing OpenAI SDK code ports with a base-URL swap. A single response defaults to 131k output tokens and can be raised to the full 1M, streaming returns reasoning and answer text as separate deltas, and structured output uses strict json_schema. Tool calling, tool_choice: required, and dynamic tool loading all work through the standard request. On the platform K3 runs text in, text out today. You can also try it in the Kimi K3 API playground first.
• Moonshot's own Kimi API. K3 is also live as model id kimi-k3 on Moonshot's OpenAI-compatible endpoint, with the same 131k default output and separate reasoning deltas.
• The Kimi apps. K3 runs in the Kimi assistant, Kimi Code, and the agent products, with max thinking effort by default.
• Self-hosting. The weights have been on Hugging Face since July 27, 2026, so you can run K3 on your own hardware, but plan for a multi-node deployment; a 2.8T MoE does not fit a workstation.

A first call looks like this:

curl -X POST https://api.unifically.com/v1/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -d '{
    "model": "moonshotai/kimi-k3",
    "messages": [
      {"role": "user", "content": "Plan a migration from Express to Fastify for this service and list the risks in order."}
    ]
  }'

The same call in Python is a base-URL swap on the OpenAI SDK:

from openai import OpenAI

client = OpenAI(
    base_url="https://api.unifically.com/v1",
    api_key="YOUR_API_KEY",
)

response = client.chat.completions.create(
    model="moonshotai/kimi-k3",
    messages=[{"role": "user", "content": "Plan a migration from Express to Fastify and list the risks in order."}],
)
print(response.choices[0].message.content)

Kimi K3 FAQ

Does Kimi K3 have an API?

Yes. Kimi K3 is live on UnificAlly as moonshotai/kimi-k3 through the OpenAI-compatible /v1/chat/completions endpoint, and on Moonshot's own Kimi API as kimi-k3. Existing OpenAI SDK code works with a base-URL swap.

How much does the Kimi K3 API cost?

On UnificAlly, $2.25 per 1M input tokens, $11.25 per 1M output tokens, and $0.225 per 1M cached input tokens. Reasoning tokens bill as output, and K3 reasons at max effort on every request.

Is the Kimi K3 API free?

No. K3 bills per token with no subscription. A typical agent step of 50,000 input tokens and 5,000 output tokens costs about $0.1125 in plus $0.05625 out, so small experiments cost cents.

How do I get a Kimi API key?

Two routes. Moonshot issues keys for its own Kimi API on the Moonshot platform. Or create a UnificAlly account (new accounts start with $0.50 of balance) and use one key for moonshotai/kimi-k3 plus every other language, image, and video model on the platform, with pay-per-token billing and no subscription.

Are the Kimi K3 weights open?

Yes. Moonshot published the full weights on Hugging Face as moonshotai/Kimi-K3 on July 27, 2026, the date it had promised, with a technical report on GitHub. They are released under the Kimi K3 License, which allows commercial use and fine-tuning but asks model-as-a-service businesses with more than 20 million US dollars in yearly revenue to sign a separate agreement.

Can I lower Kimi K3's reasoning effort?

Yes, on Moonshot's API. K3 always thinks, but reasoning_effort now accepts low, high, and max. At launch only max was available, and it is still the default when the field is left out.

Which model should you use?

K3's launch profile is long coding runs, web research, and document work, and its 42.0 on SWE Marathon plus 91.2 on BrowseComp make it the pick when the run is long, research-heavy, or output-heavy. If you want the strongest model on the depth tests instead, Claude Fable 5 leads the Arena Text board at #1, wins the knowledge-heavy benchmarks, and runs adaptive thinking rather than a fixed effort setting. GPT 5.6 Sol edges K3 on Terminal Bench 2.1, runs far faster, and has a finer effort dial with more levels than K3's three. The full head-to-heads: GPT 5.6 Sol vs Kimi K3 and Claude Fable 5 vs Kimi K3. All three run through the same endpoint, alongside every other language model on the platform.

Weighing K3 against models outside the flagship trio? We keep current per-token tables for the whole Claude lineup in the Claude API pricing guide, and the DeepSeek API breakdown covers the other budget-tier contender, V4, including its peak and off-peak windows.

We last checked this post against Hugging Face, Moonshot's API docs, and the Arena board on October 12, 2026. We will update it when Moonshot ships a new K3 checkpoint, when more independent benchmarks run K3 on shared harnesses, and when its Arena rank moves.