---
title: 'In the News: October 24, 2026 (Extra 3)'
description: 'Kimi K3.1 Pro puts a 1M-context open-weight coding model on a November 4 weights date, and its makers say they stripped cyber data from training.'
canonical_url: 'https://darkfactory.dev/news/2026-10-24-extra-3'
markdown_url: 'https://darkfactory.dev/news/2026-10-24-extra-3.md'
collection: news
date_published: '2026-10-24T23:10:00-04:00'
date_modified: '2026-10-24T23:10:00-04:00'
---

# In the News: October 24, 2026 (Extra 3)


Moonshot AI released Kimi K3.1 Pro, an open-weight coding model with a 1-million-token context window, and promised the weights by November 4. Separately, Dan Luu argues that agents have made a class of projects cheap enough that most programmers still price them wrongly.

## 1. Kimi K3.1 Pro ships now, with open weights promised by November 4

**[Kimi K3.1 Pro Tech Blog: Advancing Open Intelligence](https://www.kimi.ai/blog/kimi-k3-1-pro)** · Moonshot AI (Kimi) · Company blog, October 23, 2026

Moonshot calls K3.1 Pro its most capable "3T-class" open-weight model, with a 1-million-token context window, and says a Highspeed variant runs 2-3 times faster. Both are available today through the Kimi API and Kimi Code; the full weights are due "by Nov 4". On Terminal-Bench 4.0, run with the Claude Code harness, Moonshot reports 65.7%, against 58.2% for GPT-6 Astra and 66.4% for Claude Opus 5.5. On its own long-horizon coding benchmark it reports 61.2%, up from 55.95% for K3. All figures are the vendor's, and the competitor scores are copied from other published sources, not rerun.

The post also describes a training-time control. Moonshot says it found vulnerability-research skill rising with post-training scale, traced it to cyber-related data in the mid-training corpus and the reinforcement learning environments, and removed that data. On its in-house xdev-bench-hard exploit tasks, the checkpoint without the removal solves 90.5% at pass@3 and the released model solves 81.0%. It states the aim "not to release open weights that would extend the absolute frontier of cyber capability." Those numbers come from a benchmark only Moonshot has run.

**Why it matters:** If the weights arrive on schedule, teams that run agent harnesses on their own infrastructure get another long-context candidate to test, and Moonshot's own results are for the Claude Code harness, so a harness swap is plausible. Until the weights are out and someone outside Moonshot reproduces the coding and cyber numbers, treat them as claims.

## 2. Dan Luu says agents changed the cost of building, and most people still price it the old way

**[We should be more ambitious with the software we build](https://danluu.com/ambitious/)** · Dan Luu · danluu.com

Luu says he built a regex engine, a ripgrep fork and a faster Rust build setup for his own use, each in what he calls "a low single-digit number of hours" of his time. He estimates the same work would have taken a year or more before agents. He credits the idea to Michael Malis, who works on pgrust, a Postgres rewrite that Luu says has already surfaced real Postgres bugs. Luu's explanation for why people don't try such projects is that they carry a "pre-agentic cost model".

He is open about the limits. He did a quick comparison instead of rigorous benchmarks, so the speedups are his own unbenchmarked reports. He also writes that each more capable model leaves him further from "it just works", because what a model can implement grows faster than what it can verify without handholding. He includes a dissent from Yossi Kreinin, who argues most people should build less with AI, since many problems have no clear pass/fail criteria for an agent loop to use.

**Why it matters:** The argument is about cost estimates, and the evidence is one practitioner's projects. Still, the verification gap he describes is the part to check against your own work before scoping something large.