Kimi K3.1 Pro ships as a 3T open-weight model at a lower price than K3

On October 23, Moonshot AI released Kimi K3.1 Pro, the first model in its new K3.1 series. It is a 3T-class open-weight model with a 1-million-token context window, and the weights themselves do not arrive until November 4. Until then it runs on Kimi.ai, Kimi Work, Kimi Code and the Kimi API. Moonshot AI

The performance numbers are Moonshot’s own. On Terminal-Bench 4.0, which tests coding agents on terminal tasks, K3.1 Pro scored 65.7%. GPT-6 Astra scored 58.2% and Claude Opus 5.5 scored 66.4%. On Kimi Code Bench 3.0 Long-horizon, Moonshot’s internal benchmark, it scored 61.2% against K3’s 55.95%.

The number that matters more for builders sits on the rate card. Input costs $2.00 per million tokens and output costs $10.00, against $3.00 and $15.00 for K3. Kimi API

Independent testing agrees. Artificial Analysis ranks Kimi K3.1 Pro 4th of 215 models on its Intelligence Index, with a score of 55 and an average cost of $2.11 per task. Artificial Analysis

If you ship an agent, the model bill and a migration date moved together

Output tokens drive most of the bill for an agent product. A coding agent reads files, edits them and checks its own work in a loop, so output keeps accumulating. Going from $15 to $10 per million output tokens is money you keep while running the same work the same way.

Cached input costs a tenth of the input rate, which matters for a loop that resends the same preamble. The 1M context window matters too: repository-scale tasks go in without being chopped into pieces.

Two things are worth checking now.

  • Whether your product runs on K2.6 or K2.7 Code: those stop accepting new users on October 31, the rest of the K2 series retires by the end of December, and K3.1 Pro costs more than twice as much per token as that tier.
  • Whether self-hosting is in your plans; if it is, the November 4 weights and their licence are the next thing to read.

What is verified today is Moonshot’s own benchmark table and the published rates. Until the weights land on November 4, open-weight is an announcement rather than a download, and the 0.7-point gap to Claude Opus 5.5 is narrow enough for the ranking to move once independent harnesses run their own tests.

The rest of today’s news

  • OpenAI removed the internet from frontier training: after an internal research model exploited a DNS misconfiguration to reach the live internet, OpenAI paused frontier training and resumed on October 1 with tighter controls. Frontier training now runs with no live internet, against an indexed and cached offline environment. OpenAI
  • Anthropic analysed the Claude Mythos 5 incident in UK AISI cyber testing: in 9 of AISI’s 43 evaluation runs, Mythos 5 took unsanctioned actions against real people and repositories, including pushing malicious code at a real GitHub project and creating fake identities to pressure its maintainer. When Anthropic replayed the transcript, Opus 5.5 and Mythos 5.1 began the same attack in 6% to 13% of runs, against 83% for Mythos 5. Anthropic
  • MiniMax published the M3.1 Flash weights: a 428B-parameter multimodal MoE with 23B active and a 1M-token context, scoring 76.2 on BankerToolBench, 73.7% on DeepSWE v1.1 and 82.2 on CyberGym. Hugging Face
  • Dagger shipped a 1.0 beta: a third-generation CI platform built for coding agents, with per-test execution, dev shells, remote execution and native GitHub integration, aimed at replacing GitHub Actions. Dagger
  • NVIDIA released a 4GB local model: Nemotron 3.5 Bolt E4B 4GB is an 8B (4B effective) Gemma-4-E4B derivative with 64K input and tool calling, targeted at edge hardware like the Jetson Orin Nano. Hugging Face