🛠️ Development

Everything on the development side — agents, testing, models, tooling, and how teams build with AI. Sorted into here by topic, not by which Slack channel it came from.

BlogArtificial Analysis

GPT-6 Astra's Early Coding-Agent Benchmarks

Artificial Analysis's independent read on GPT-6 Astra's early benchmark numbers against existing frontier models.

Scores roughly equal to Claude Fable 5 on the Coding Agent Index, but at a lower cost per task.
Uses fewer tokens than GPT-5.6 Sol for similar Intelligence Index performance, though that efficiency is outweighed by pricing 2.5x higher than GPT-5.6.

View ↗

added by Radar • 4th Sep 2026

BlogOpenAI

GPT-6 Astra: A New Generation of Intelligence

OpenAI's announcement of GPT-6 Astra, its new flagship model, which it bills as state-of-the-art on computer use, browsing, software engineering, cybersecurity, science, and professional work.

Rolled out first to a limited set of organizations, then to all ChatGPT Plus, Pro, Business, and Enterprise users, plus the OpenAI API and AWS.
The launch itself was rocky — several outlets published coverage citing OpenAI's material before the official Astra page was publicly live, and OpenAI briefly pulled the post.

View ↗

added by Seb Kay • 3rd Sep 2026

DocsGitHub

GitHub CLI: media in issues, pull requests, and comments

GitHub CLI's new repeatable `--attach` flag, which uploads a local image or video and references it inline in an issue, pull request, or comment body.

Works on `gh issue create/edit/comment` and `gh pr create/edit/comment`; supports PNG, JPEG, GIF, WebP, SVG, MP4, MOV, and WebM.
Size limits match the web upload flow — 10MB for images/GIFs, 10MB for video on Free plans, 100MB on paid plans; requires gh CLI 2.99.0+ and repo write access.
Repeatable per command, and `--attach './image.png#Description'` sets alt text.

View ↗

added by Adam Tomat • 2nd Sep 2026

BlogSean Goedecke

LLMs Reward Expertise

An argument that prompting well isn't a generic, democratized skill — deep domain expertise is what lets someone steer an LLM effectively.

Uses Terence Tao's math conversations with ChatGPT as an example: his edge comes from spotting flaws and better approaches, not from prompting technique.
The limiting factor in AI collaboration is communicating exactly what kind of solution you want and judging the output — something only domain expertise enables.
As models improve, the human stays the bottleneck for complex, nuanced tasks, so domain expertise keeps its value rather than being flattened by AI.

View ↗

added by Adam Tomat • 2nd Sep 2026

BlogAnthropic

Introducing Claude Fable 5.1 and Claude Mythos 5.1

Anthropic's announcement of Claude Fable 5.1 and Claude Mythos 5.1: the same underlying model at two safeguard levels, Fable generally available and Mythos restricted to vetted cybersecurity/life-sciences users.

Fable 5.1 scores 55.8% on Terminal-Bench 4.0 agentic coding tasks, up from Fable 5's 42.0%, and cache-read pricing drops 75% to $0.25/M tokens — cutting typical workload cost ~25% and agentic-task cost up to 45%.
Mythos 5.1 is aimed at scientific research use, credited with designing protein binders with 10x higher binding affinity than prior winners and a Venus elevation map at 2-3km resolution (up from 10-20km).
Safeguard false positives in cybersecurity are down 60%, and the model now permits vulnerability discovery (though not exploit development).

View ↗

added by Seb Kay • 1st Sep 2026

BlogLimestone Digital

AI Velocity Pods: One Engineer Plus Agents vs. a Traditional Squad

Limestone Digital's description of its "AI Velocity Pod" model for PE-backed portfolio companies: one dedicated senior engineer paired with AI agents, backed by a lean, shared support team, running inside the client's existing codebase and standups.

A pod is roughly 1.5 FTE total — one full-time forward-deployed engineer plus pooled AI-productivity, PM/BA, CSM, and AI-QA support — billed as one monthly price rather than time-and-materials.
Progress is tracked with an in-house "Dev Intelligence Platform" against each client's own baseline (cycle time, PR velocity, change-fail rate, MTTR, AI-assist hours) rather than industry benchmarks.
Case studies claim 42% faster delivery, 85% fewer incidents, and 35% faster cycles with 40% higher predictability — though these are self-reported, baseline-relative figures, not third-party verified.

View ↗

added by Radar • 29th Aug 2026

BlogScott M. Graffius

Meta's CTO Called Their AI Reorg 'Atrocious' — Lessons for Human-AI Teams

A case study of Meta's "Project OT" AI-native restructuring — which cut roughly 8,000 jobs aiming for smaller, AI-leveraged teams — arguing it stumbled on people and team dynamics rather than technology.

Employee morale dropped from 74% to 55% favorable, major technical incidents rose 40%, and firefighting time was up 70% — even as AI-assisted code changes rose 220% while user-facing features grew only 36%.
Meta CTO Andrew Bosworth admitted the company did an "atrocious" job explaining the Applied AI vision and support during the transition; a planned second restructuring wave was cancelled.
Draws five lessons on "exotic team dynamics": smaller teams need new roles and workflows (not just fewer people), accountability blurs between humans and AI, trust needs to be architected in, activity isn't productivity, and org charts lag how human-AI teams actually work.

View ↗

added by Radar • 29th Aug 2026

BlogOpenAI

OpenAI Ends Its Partnership With Cursor After SpaceX Acquisition

OpenAI's announcement that it is cutting Cursor's direct access to its models, effective November 12, after SpaceX's acquisition of Cursor.

OpenAI cites Elon Musk–linked companies' history of violating its terms of service as the reason it can't be confident SpaceX will keep using its models within contract terms.
OpenAI models reportedly account for only around 5% of Cursor's traffic; users can still bring their own OpenAI API key, and OpenAI's own IDE extensions remain available.

View ↗

added by Seb Kay • 29th Aug 2026

🛠️ Development ​

🛠️ Development