Kimi K3.1 Pro Pricing: Cost per Task and Monthly Bill

Last updated: October 2026

Key takeaways

• Standard API pricing: $2 input, $10 output, $0.20 cached input per 1M tokens, versus K3 at $3, $15 and $0.30.
• Pro Highspeed is reported at twice standard pricing for 2 to 3x faster output.
• Artificial Analysis puts cost per task at $1.98 on its Intelligence Index, with tokens per task up from 48K to 55K.
• In an illustrative 50k-conversation support bot, the monthly bill falls from $900 on K3 to $600 on K3.1 Pro, and rises to $1,200 on Highspeed.
• Weights were not available at launch. Moonshot promised an open-weight release for November 4, 2026.

What is Kimi K3.1 Pro?

Per AINews on Latent Space, Kimi K3.1 Pro is a multimodal mixture-of-experts model with 2.8T total parameters and 104B active, plus a 1M-token context window. It is positioned as roughly matching the top closed model class and narrowing the lead to about a month. The source calls it Kimi 3.1 Pro; the vendor naming is K3.1 Pro.

For a pricing page, the interesting part is the rate card and what it does to cost per task, not the leaderboard. You can see how Moonshot's other models are listed on the Moonshot pricing page on Calcaas.

How much does Kimi K3.1 Pro cost?

| Model or tier | Input per 1M | Output per 1M | Cached input per 1M | |---|---|---|---| | Kimi K3 (previous) | $3 | $15 | $0.30 | | Kimi K3.1 Pro (standard) | $2 | $10 | $0.20 | | Kimi K3.1 Pro Highspeed (derived at 2x, not a published rate card) | about $4 | about $20 | about $0.40 |

The first two rows are reported figures. The Highspeed row is our estimate from the reported "twice standard pricing"; the source gives no separate rates, so check Moonshot's price list.

Is the price cut real once you count tokens per task?

Per-token price is not per-task price. Artificial Analysis reports cost per task of $1.98 on its Intelligence Index, $1.23 median on Arena WebDev and $2.81 on its Cyber Index. On the Intelligence Index, output grew to 55K tokens per task from K3's 48K, about 15% more.

Here is the original arithmetic. The price per token fell to two thirds of K3's (0.67). Tokens per task rose by a factor of about 1.15. Multiply them and you get about 0.76, so the effective saving per task is roughly 24%, not 33%. That is illustrative, because it assumes your token mix moves like the benchmark's, but it is a better planning number than the headline.

How does it rank against closed models?

On the Artificial Analysis Intelligence Index, the source reports Kimi K3.1 Pro at 55, fourth place. Opus 5.5 scored 58, and Sonnet 5.5 and Muse Wonder Full scored 56. It also reports #8 on Arena Code WebDev at 1,692 points, and AA-Omniscience accuracy of 51% with a hallucination rate of 44%.

A 3 point gap to the top is the kind of difference that rarely changes a support bot's output but can change a hard coding agent's. Test on your own tasks. For a broader comparison of open and closed options, see open-weight versus frontier LLM cost.

What does a 50k-conversation support bot cost on K3.1 Pro?

Say a support bot handles 50,000 conversations a month, each with 3,000 input tokens and 600 output tokens. That is 150M input and 30M output tokens a month. All figures are illustrative.

| Setup | Input cost | Output cost | Monthly total | |---|---|---|---| | Kimi K3 | $450 | $450 | $900 | | Kimi K3.1 Pro (standard, no caching) | $300 | $300 | $600 | | Kimi K3.1 Pro (60% of input cached) | $138 | $300 | $438 | | Kimi K3.1 Pro Highspeed (about 2x, no caching) | $600 | $600 | $1,200 |

Highspeed doubles the bill to make responses 2 to 3 times faster. That is worth it for voice or live agents where users watch the cursor, and a poor trade for batch work.

When do open weights change the math?

The weights were not released at launch, though Moonshot promised an open-weight release for November 4, 2026, and the licence terms were not stated. If they arrive on friendly terms, self-hosting becomes a break-even question: GPU hours versus API spend. We cover that trade in Moonshot's token throughput as a market signal and the self-hosting posts it links to. Until the weights and licence are public, budget on API rates.