AI News, page 4 of 6

AI and software engineering, as it happens.

GitHub measured that shortening a tool's output made Copilot more expensive, not less

GitHub published the numbers behind its work on Copilot token cost and reported that a tool built to shorten command output raised the total cost of a task, because the model reran the command to recover what had been cut.

Productivity

The ChatGPT Codex desktop app carries 1.7 GB of bundled binaries, including a headless LibreOffice

Simon Willison found that the OpenAI Codex desktop app keeps 1.7 GB in a local cache folder, including full copies of Python, Node.js, git, Poppler and a 429.7 MB headless LibreOffice, with agent skills that tell the model how to run them.

Dev tools

Slotstream runs a 125B mixture-of-experts model on a normal Mac by streaming from SSD

Slotstream is a new MIT-licensed Swift binary that runs Qwen3.8-Flash-Next, a 125-billion-parameter model whose weights are 105 GB on disk, on an 8 GB Mac by streaming the experts from SSD and holding only what a token needs in RAM.

Open source

Cursor lets its cloud agents run on machines your team already owns

Cursor shipped self-hosted machines on 2 September, so a team can point its cloud agents at hardware it already owns instead of running every tool call on Cursor's servers.

Dev tools

GitHub Copilot deprecated five models on 1 September, including Gemini 3.1 Pro, Claude Opus 4.5, and both Claude Sonnet 4.5 and 4.6

GitHub retired Gemini 3.1 Pro, Claude Opus 4.5, Claude Sonnet 4.5, Claude Sonnet 4.6, and Raptor Mini across most Copilot surfaces on 1 September 2026, with named replacements for each.

Dev tools

Anthropic published a reference blueprint for Claude commerce agents, with shopping and merchant roles over three runtimes

Anthropic released a public reference implementation of a Claude shopping agent and a Claude merchant agent that run on the Messages API, the Claude Agent SDK, and Managed Agents, with four demo verticals and a Claude Code plugin to scaffold your own.

Models & agents

Mezmo open-sourced Aura, an SRE agent platform where the whole agent definition sits in a reviewable TOML file

The observability company Mezmo released Aura, an Apache-licensed Rust agent for investigating and fixing production incidents whose entire agent system is defined in TOML, with 267 stars on the repository.

Open source

Anthropic launched a public tool to check whether a file was made with Claude, using C2PA credentials and text watermarks

Anthropic put up a web page where anyone can upload an image, audio or video file up to 100 MB and see whether it carries a Claude-signed C2PA content credential, with a separate text-watermark detection API in private preview.

Models & agents

Zoetrope turns a Claude Code session into a live flow graph you can scrub through

Zoetrope is a new MIT-licensed local viewer that reads Claude Code session transcripts and renders the agent, its subagents, and every tool call as an interactive graph, and the repository has picked up 766 stars.

Dev tools

One developer rewrote a 65,000-line Go terminal editor in Rust with Claude Fable for about $400

Iurii Krasnoshchok reports that he rewrote his 65,000-line Go terminal editor "rune" in Rust using Claude Fable, for a token cost of about $400, by treating the code as data and orchestrating subagents rather than editing files directly.

Productivity

FrontierHarness Eval finds cost per task varies 17 times across nine agent harnesses running the same model

A public leaderboard from Runta ran the same model, Kimi K3, through nine coding agent harnesses and reports that the median cost per task ranges from $1.05 to $18.34, before pass rates are considered.

Dev tools

Vercel published the design brief its coding agents load before they build a page

Vercel published design.md, a single public file that its coding agents load before building a page, and its evaluation shows pages built with it had 57 percent fewer failures than pages built without.

Productivity

A third of Perplexity's cited pages do not contain the number they are cited for, says a 1,826-citation audit

A Haus Research audit of Perplexity's sonar and sonar-pro models found that 34.7 percent of citations attached to a number pointed at a page that would not open or did not contain the number, across 210 technology companies and ten question templates.

Go-to-market

Google ships Gemini 3.8 Flash at $0.75 per million input tokens and puts it into Google Search AI Mode

Google announced Gemini 3.8 Flash on 2 September at introductory pricing of $0.75 per million input tokens and $3.75 per million output tokens, with a separate 3.8 Flash Cyber variant for vetted security work.

Models & agents

Three brand-new sites produced 215,128 "best software" pages, and Perplexity cited them 181 times

A Trellner Research report says three coordinated sites created 215,128 template pages for "best of" software categories, and Perplexity cited them 181 times across 41 categories in its sample.

Go-to-market

An AI security tool found six curl CVEs after Anthropic and OpenAI's tools found none

A startup called AISLE ran its own AI security tool against curl and got six CVEs assigned, after Anthropic and OpenAI's security tools were reported to have found zero in the same code the week before.

Dev tools

Airship puts a Figma-style canvas in front of a running dev server and lets Claude Code or Codex edit the source

Airship is an MIT-licensed CLI that opens a Figma-style canvas in front of a running dev server, so a developer can click any element and ask Claude Code, Codex or OpenCode to change the source.

Dev tools

Moli, a Rust headless browser built for AI agents, skips layout and paint on structure-only work

Moli is a Rust-built headless browser aimed at AI agents that generates layout and pixels only on request, so a page fetch that never needs a screenshot skips the paint stage entirely.

Open source

HarnessRouter runs Codex, Claude Code and DeepSeek Harness from one self-hosted Docker container

HarnessRouter Community Edition is a self-hosted Apache-2.0 container that runs Claude Code, Codex, Hermes, Pi and DeepSeek Harness through one API, with a shared console and per-session Linux users for isolation.

Open source

Open-source SEO and GEO skills for Claude wire keyword research to real GSC, GA4 and Ads data

A new MIT-licensed set of eight SEO and GEO skills for Claude runs keyword research, rank tracking and AI-visibility checks against a site's own Search Console, GA4 and Google Ads data.

Open source

Marketing OS bundles fourteen marketing modules, including a GEO audit, into one Claude skill

Marketing OS is a single MIT-licensed Claude skill that carries fourteen marketing modules, from a scored site audit to a GEO citability check and a hook-generation engine.

Open source

Alibaba's AMAP-ML releases LongHorizon-Harness, a manager-executor-auditor loop for computer-use agents

Alibaba's AMAP machine learning team open-sourced a harness that wraps existing coding agents in a Manager-Executor-Auditor loop for long-running computer-use tasks.

Open source

ApodexAI open-sources FrontierAgent, a Python agent framework that reached 1,360 stars in 10 days

ApodexAI released FrontierAgent, an Apache-2.0 Python agent framework with a terminal UI and a coordinator-plus-subagents mode, alongside its Apodex-1.1 model.

Open source

shadcn ships an official chatbot template, 889 stars in three weeks, built on new shadcn/react and shadcn/typeset packages

shadcn published a minimal official chatbot template on Next.js and the AI SDK that shows off two newer packages, shadcn/react and shadcn/typeset, and it collected 889 GitHub stars in its first three weeks.

Dev tools

GitHub Copilot code review can now approve pull requests, off by default across three admin levels

GitHub Copilot code review can now submit approvals that count toward a repository's merge requirements, and admins decide whether to turn that on at three levels.

Dev tools

Vercel Labs ships fx, a 7.8 MiB coding agent written in Zig with 2,664 stars in three weeks

Vercel Labs open-sourced fx, a compact coding agent harness written in Zig that ships as a 7.8 MiB binary and is designed to be embedded inside other software.

Dev tools

Gemini adds agentic video understanding that cuts analysis cost by up to 66 percent and token use by up to 88 percent

Google turned on an agentic video mode for Gemini 3.7 Flash, 3.6 Flash and 3.5 Flash-Lite that decides which frames, audio segments and transcripts to read, with Google reporting up to 66 percent lower analysis cost, up to 88 percent lower token use and up to 7 percent higher accuracy.

Models & agents

Anthropic ships Claude Fable 5.1 and Mythos 5.1 with a 25 percent lower price and a 75 percent cheaper cache read

Anthropic released Claude Fable 5.1 for coding and knowledge work, with input at $10 and output at $50 per million tokens, cache reads at $0.25 per million, and immediate availability on Amazon Web Services, Google Cloud and Microsoft Azure.

Models & agents

An MCP server for x64dbg exposes 84 debugger tools to any AI assistant, and picks up 1,824 stars in 10 days

A new Zig plugin makes x64dbg controllable by any MCP-compatible AI assistant, exposing 84 debugger tools including breakpoints, memory reads, register dumps and pattern scans over authenticated HTTP.

Open source

Cumora is a team chat where AI agents live in the same channels as humans, and it has 3,385 stars in 15 days

Cumora is a fifteen-day-old open source team chat where AI agents sit in the same DMs and group conversations as people, hold their own personas and memory, and can send and receive real email.

Open source

Archive

179 items published so far, by month.

All months