---
title: 'In the News: November 18, 2026 (Extra)'
description: 'Kernel traces why Gemini 4 Carbon takes close to an hour on bug fixes; Microsoft prices MAI Code 2 for everyday Copilot work.'
canonical_url: 'https://darkfactory.dev/news/2026-11-18-extra'
markdown_url: 'https://darkfactory.dev/news/2026-11-18-extra.md'
collection: news
date_published: '2026-11-18T10:06:00-05:00'
date_modified: '2026-11-18T10:06:00-05:00'
---

# In the News: November 18, 2026 (Extra)


Kernel traced every step of Gemini 4 Carbon's runs on a private bug-fixing eval and found that about half of them went on work the ticket never asked for. A check-in injected by the harness shortened one trial by nine minutes. Microsoft has priced two new coding models for routine Copilot work and says plainly they are not tuned for long agentic runs. An NBER paper from the spring measures how much of the extra code agents write turns into releases.

## 1. Kernel finds Gemini 4 Carbon spends half its steps on work nobody asked for

**[Why is Gemini 4 Carbon taking so long?](https://kernel.dev/blog/why-is-gemini-4-carbon-taking-so-long/)** · Kernel · kernel.dev, November 2026

On one of its private evals, Kernel gives an agent a codebase from just before a bug was fixed and a report that describes the bug only by its symptoms. Carbon took close to an hour on tasks that Opus 5.5 and GPT-6 Astra finished in minutes. Slower inference explains part of the gap: 16 to 30% of Carbon's extra time against Astra, and 45 to 49% against Opus. The rest is longer trajectories. Against Astra, Carbon took about seven times as many steps and eleven times as many actions, and processed more than sixteen times as much input.

Kernel labelled each step. Of Carbon's 215 steps in the labelled runs, 117 were waste: 70 auditing code beyond the bug and 47 closing loopholes the ticket had already ruled out. Astra wasted one step of 19. Carbon also re-read code: 39% of the lines it read were lines it had already seen, against 7% for Astra. Running Carbon in opencode and mini-SWE-agent made no meaningful difference, and on low effort it often failed to produce a fix at all.

To test whether this was a knowledge gap or a habit, Kernel re-ran one trial with the harness injecting "do you already have enough to make the fix?" every five turns. That trial finished in 28 minutes instead of 37, and the wait between naming the cause and the first edit fell from about 25 minutes to about 10. The eval is private and the check-in result comes from a single sample trial. The post was at 200 points and 100 comments on Hacker News at 10:02 EST today.

**Why it matters:** On Kernel's numbers, most of Carbon's extra time against Astra is extra steps, so faster inference alone would close only part of the gap. A periodic "are you ready to edit?" prompt is a harness change you can test on your own tasks before you switch models.

## 2. Microsoft prices MAI Code 2 for everyday Copilot work, not long agent runs

**[MAI Code 2 and Code 2 Flash: Everyday coding at scalable prices](https://microsoft.ai/news/introducing-mai-code-2-and-mai-code-2-flash/)** · Microsoft AI · microsoft.ai, November 17, 2026

Microsoft AI released two coding models at Ignite. It says that averaged over DeepSWE 1.1, Terminal Bench 2.1, SWE-Atlas and FrontierCode v1.1, MAI Code 2 gives 98% of GPT 6.1 Sol's performance at 40% of the cost, and 96% of Claude Opus 5.5's at 5% of the cost. Code 2 Flash is aimed at high-volume work such as targeted edits and routine fixes. Microsoft is direct about the limit: the models are "explicitly not yet tuned for harder problems, like fundamental science research and very long horizon agentic work."

All of these figures are Microsoft's. Three of the four benchmarks ran on MSBench, Microsoft's internal eval platform, through the GitHub Copilot CLI without subagents; FrontierCode scores and costs come from Artificial Analysis. Both models' costs are shown at a 50% promotional price. Both are available in GitHub Copilot now, Code 2 Flash is in limited preview in Microsoft Foundry, and Microsoft says OpenRouter, Vercel, Hermes, Open Code and Kilo are coming.

**Why it matters:** This gives you a cheap default for the short, routine edits that make up much of a factory's volume. By Microsoft's own description it is the wrong model for long autonomous runs, and the cost comparison will move when the promotional price ends.

## 3. NBER paper: agent commits rise 240%, releases 30%

**[Writing Code vs. Shipping Code: Productivity Effects Across Generations of AI Coding Tools](https://www.nber.org/papers/w35275)** · Mert Demirer, Leon Musolff and Liyuan Yang · NBER Working Paper 35275, May 2026, revised September 2026

The paper combines data on more than 500,000 GitHub developers with their AI usage telemetry. It finds cumulative increases in commits of 30% for autocomplete, 180% for interactive coding agents and 240% for autonomous coding agents. The 240% falls to 80% for the number of projects and to 30% for actual releases. The authors read this as the weak-link hypothesis, with an estimated elasticity of substitution of 0.23 between AI and human effort. Across four software marketplaces they find a sharp increase in new apps but no increase in total usage.

This is a working paper, and we read the abstract and disclosures rather than the full PDF. NBER's disclosures say Demirer and Musolff each previously held a postdoctoral position at Microsoft and now work as paid research consultants for the company.

**Why it matters:** Commit counts are the wrong scoreboard for an agent rollout. On these figures most of the extra output stops before release, so measure releases and usage, and look hard at the human steps after the commit.
