---
type: analysis
slug: kimi-k3-1-pro-s-token-price-is-not-its-task-price
title: "Kimi K3.1 Pro’s token price is not its task price"
description: "Moonshot’s new model pairs a million-token context with a steeply higher output rate. Artificial Analysis’s reported output use is a reason to measure full task cost—not a prediction of what any one buyer will pay."
published_at: 2026-10-26T17:00:00.000Z
topics: ["trends"]
source_urls: ["https://simonwillison.net/2026/Oct/23/kimi-k31-pro","https://vercel.com/changelog/kimi-k3-1-pro-and-kimi-k3-1-pro-highspeed-are-available-now-in-ai-gateway","https://huggingface.co/moonshotai/Kimi-K2-Thinking"]
author: "Andrey Reshetnikov"
author_profile: https://newruntime.com/owner-profile.md
---
# Kimi K3.1 Pro’s token price is not its task price

Moonshot’s new model pairs a million-token context with a steeply higher output rate. Artificial Analysis’s reported output use is a reason to measure full task cost—not a prediction of what any one buyer will pay.

## Summary

Moonshot has released Kimi K3.1 Pro through its API, describing it as a mixture-of-experts model with 2.8 trillion total parameters, 104 billion active parameters, and a context window of up to one million tokens. Standard pricing is $2 per million input tokens and $10 per million output tokens; a Highspeed tier is described as two to three times faster and costs twice the standard rate, except for input cache hits. Vercel has added both versions to AI Gateway. Moonshot says full model weights are planned for November 4, 2026.

## What happened

The launch offers a hosted model with low, medium, and max reasoning settings, a very large context window, and a faster paid tier. Moonshot and Vercel cite improvements in areas including coding, 3D modeling, financial analysis, and agentic tasks, but those claims do not establish performance on a buyer’s workload. The model is not yet available as open weights. Artificial Analysis reports an index task cost of $2.11 and unusually high output-token use; those are measurements on its index, not a quote for a customer’s application. Together, the launch pricing and that output-use signal make total task cost a more useful question than input price alone.

## What the archive adds

The archive includes a fragment on Moonshot’s earlier Kimi K2 Thinking, described as an open-weight reasoning agent with a 256K context window and long tool-use chains. That is relevant company context, but it cannot establish K3.1 Pro’s capabilities or whether its promised weights will arrive on schedule. The supplied research dossier also identifies an article framing Kimi K3 as an open-weights escalation, but provides no findings from it that would support stronger claims about the current launch. The useful precedent here is limited: hosted access and an eventual weight release are distinct stages, and the second should not be treated as available today.

## Trend signal

Model launches are packaged around more than capability claims: context limits, latency tiers, API access, and future deployment options all influence adoption. But list prices are only inputs to an economic comparison. Long answers, retries, and review can alter the cost of completing a task, while a faster tier matters only if response time improves the workflow. The inference is that buyers should evaluate the model by cost per accepted result at their own quality and latency thresholds, not by token rates or maximum context in isolation.

## Why now

According to Moonshot’s launch materials, K3.1 Pro is API-accessible before the promised full-weight release, so the current product is hosted access rather than self-hosting. Moonshot’s listed input and output rates differ substantially, making output volume a concrete cost driver. Vercel’s AI Gateway listing adds another route to access the hosted model. Artificial Analysis reports unusually high output use on its index, which makes measuring generated tokens particularly relevant, without establishing a buyer’s workload cost.

## Connections to other events

The key number to scrutinize is not the one-million-token ceiling but the relationship between input and output charges. Standard output tokens cost five times as much as input tokens per million, and Highspeed doubles the standard rate apart from input cache hits. Artificial Analysis’s reported $2.11 index task cost and unusually high output use reinforce the need to track generated tokens. That is a warning to measure, not evidence that a particular customer will face the same bill: its index task may differ from a real workload in prompts, output length, retries, and acceptance criteria.

## Why it matters

Builders can try K3.1 Pro through the API and Vercel AI Gateway without waiting for weights, but should treat a pilot as an evaluation rather than proof of fit. Compare standard and Highspeed with a current baseline on representative tasks, counting output tokens, accepted results, retries, review effort, and latency. A larger context window may enable useful experiments with long documents; it does not mean every document should be sent in full. Buyers should make the decision on completed work and operating constraints, not launch-scale specifications.

## What happens next

- **At least one third-party evaluation will report K3.1 Pro output use or cost on a named workload, clarifying how its index measurements compare with practical tasks.** (6 months, 65%). Watch for: A published independent evaluation gives token use or spend for defined coding, long-context, or agent tasks.
- **Moonshot will either make the promised K3.1 Pro weights available by its stated target or publicly revise the target.** (12 months, 60%). Watch for: A dated release or repository announcement appears, or Moonshot explicitly changes the November 4, 2026 target.

## Limits and uncertainty

The primary evidence does not establish independent performance for a particular workload, deployment, or latency target. Artificial Analysis’s index cost and output-use observation are not a forecast of an application’s bill. Nor do the supplied materials establish whether a million-token prompt improves results enough to justify its cost, or whether Highspeed’s stated acceleration holds for a buyer’s traffic. The planned weight release is a future commitment, not a present self-hosting option.

## Skill to build now

Build a small cost-per-accepted-result test this week: select representative tasks, define acceptance criteria in advance, and compare K3.1 Pro Standard and Highspeed with your current baseline. Log output tokens, retries, review time, cost, and latency.

## Reader takeaway

This week, run one focused bake-off: compare K3.1 Pro Standard and Highspeed against your current baseline on representative tasks. Use the same success rubric and record output tokens, cost, accepted results, retries, review time, and latency. Decide from cost per accepted result; do not extrapolate your bill from Artificial Analysis’s index task.

## Open question

Will K3.1 Pro’s unusually high output use on Artificial Analysis’s index persist on real coding and agent workflows, or do those tasks produce a materially different cost profile?

## Sources

- [Moonshot released Kimi K3.1 Pro , a 2.8T total / 104B active MoE with 1M-token context.](https://simonwillison.net/2026/Oct/23/kimi-k31-pro)

- [Kimi K3.1 Pro is live on AI Gateway in regular and highspeed versions, with 1M context and gains in coding, 3D modeling, and finance.](https://vercel.com/changelog/kimi-k3-1-pro-and-kimi-k3-1-pro-highspeed-are-available-now-in-ai-gateway)

- [huggingface.co](https://huggingface.co/moonshotai/Kimi-K2-Thinking)

- [Kimi K3.1 Pro](https://www.kimi.ai/blog/kimi-k3-1-pro)

- [Kimi API Platform](https://platform.kimi.ai/)

- [Kimi K3.1 Pro — Artificial Analysis Intelligence, Performance & Price Analysis](https://artificialanalysis.ai/models/kimi-k3-1-pro)

- [Kimi K3.1 Pro in Artificial Analysis](https://artificialanalysis.ai/pt/articles/kimi-k3-1-pro-launch-benchmarks)

- [Kimi K3.1 Pro Benchmark Results for Data Analysis](https://plotly.com/blog/kimi-k3-1-pro-plotly-data-analytics-bench/)

- [Kimi K3.1 Pro: API Provider Performance Benchmarking & Price Analysis](https://artificialanalysis.ai/models/kimi-k3-1-pro/providers)

Author: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- topic hub: [Trends](https://newruntime.com/topics/trends.md): Explore the trends topic hub.
- collection: [Analysis](https://newruntime.com/analysis.json): Open the latest authorial analysis.
- collection: [Source Ledger](https://newruntime.com/source-ledger.json): Inspect the public source evidence graph.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)