Newsletter 047 - September (week 39)
Hi, and welcome to this week's edition of the newsletter!
As always, feel free to reply to this email to get in touch. Or ping me on any of my socials, which can be found on my personal website, danclarke.com.
If you get value from this newsletter, please help me out by sharing it on your socials and with your tech friends and colleagues - it really helps me spread the word and keep the newsletter going 🙏 (at-mention me, and I'll repost).
📌 News item of the week
Anthropic and OpenAI both released new models on September 22, about an hour apart, and both put the price first.
Claude Opus 5.5 is cheaper per token than Opus 5, with the biggest cut on cache reads, which Anthropic says make up most of the cost of agentic and coding work. It also uses fewer tokens per task, so it's cheaper again in practice, and Anthropic says it performs at the level of Fable 5.1 on most work. It's now the default Opus in Claude Code.
Opus 5 has had a lot of hate, especially around the way it writes responses - using very hard to read language and convoluted sentence structures. I'm seeing a lot of really good feedback on Opus 5.5, with plenty of people happy that "Claude is back!"
OpenAI's GPT-6 Sol and GPT-6 Luna also came with lower prices. OpenAI compared them with its own previous models, saying both are half the price of GPT-5.6's promotional pricing.
📰 News items and dev picks
AI news
New models
- Claude Opus 5.5 and GPT-6 Sol and Luna - My News item of the week this edition - see the full write-up at the top.
- Grok 4.7 - SpaceXAI released Grok 4.7 on a new, larger base model trained for tasks that run for hours, at the same $2/$6 per million tokens as Grok 4.6.
- MiMo-V2.6 - Xiaomi's new open-weights model is MIT licensed, with 1.02 trillion total and 42 billion active parameters in the Pro checkpoint and a 1M-token context.
Safety and escaped agents
- OpenAI agent in Medicare portal - Australia's Prime Minister said an OpenAI agent got round access blocks into non-public parts of the Medicare statistics portal on June 18, and that OpenAI didn't tell the government until September 10.
- Gemini got into three companies - Google confirmed that a Gemini model accessed the systems of three real companies during a May security test, after real internet access was unintentionally left available, and said the model stopped each time.
- Prompt injections in compaction summaries - In one of its new misalignment reports, OpenAI describes an unreleased model writing fake "BREACH ALERT" instructions into its own context-compaction summaries during training.
- ZCode uploads your Git history - A researcher found Zhipu's ZCode coding agent uploading whole workspaces, full
.githistory included, before every prompt, whatever the privacy settings said.
Coding agents and editors
- Claude Code 2.1.277 to 2.1.282 - Claude Code now reads
AGENTS.mdwhen a project has noCLAUDE.md, and auto mode for API and Enterprise users now uses a server-side classifier by default. - Claude Code cloud sessions GA - What used to be Claude Code on the web is now generally available, and existing Pro and Max subscribers can claim a one-time credit to try it ($100 on Pro, $250 on Max) with
/claim-creditby October 7. - Copilot runtime rewritten in Rust - Stephen Toub explains how mostly one developer and a set of AI agents ported the Copilot runtime from about 430,000 lines of TypeScript to 832,378 lines of Rust in around fourteen and a half weeks.
- JetBrains Air - JetBrains brought its agent products together under one name, with Air in the IDEs, Air Teams and Air Governance, and third-party agents plug in over the Agent Client Protocol.
- VS Code 1.139 - Agents can now run inside a project's Dev Container on SSH, Tunnel and WSL remote hosts, not just locally.
- Safari MCP - Safari 27 ships a local MCP server, so a coding agent can drive the browser to check rendering, styles, accessibility and load timing.
Sandboxes
- Docker Sandbox Kits - Docker is bringing its Sandbox Kit spec to the CNCF, which packages an agent, its tools and a typed list of what it's allowed to reach as a single OCI image, and on the same day launched Cloud Sandboxes on Docker-managed compute.
- Copilot app local sandboxing - The GitHub Copilot app can now, in public preview, limit each session's file, network and credential access per project, and it refuses to run if the OS can't enforce the policy.
Also from Anthropic
- Claude finds an enzyme system - About 950 Claude agents searched a DNA sequence database for 21 hours and flagged a new enzyme system with CRISPR-like repeats, though Anthropic says it doesn't yet know what it does.
- Cowork is now Claude - Claude chat and Cowork are now one app, and Claude Docs and Claude Slides have launched in beta.
- Claude Marketplace - Teams can now buy connectors, plugins and Claude-powered products from vendors like Cursor and Snowflake with part of their committed Anthropic spend.
Dotnet news
- NuGet signing certificate change - Microsoft started signing its NuGet packages with a new certificate on September 23, so if you use a trusted-signer list or
dotnet nuget verifyyou'll need to add the new fingerprint. - Uno Platform 6.7 - Uno added Android TV, Android Auto and Wear OS as targets, and its Skia renderers now only repaint the parts of the screen that changed.
- Microsoft Learn docs repos closing - Microsoft plans to retire most of the public GitHub repos behind Microsoft Learn's product docs by the end of December, though docs for open-source products and projects are expected to stay open to contributions.
- Microsoft Agent Framework update - The latest update adds AG-UI endpoints for .NET and Python, a memory provider that carries context between conversations, and CodeAct, where the agent writes a multi-step task as a single program.
Others
- TeamPCP infiltrated by Google - A Google analyst spent months undercover inside TeamPCP, the group behind a supply-chain campaign that breached more than a thousand companies, which let Google warn victims and disrupt it.
- Ubuntu kernel releases every week - Canonical is replacing its separate four-week and two-week kernel cycles with a single two-week cycle released weekly, because it says LLMs and AI agents have turned bug discovery into "a highly automated engine".
- F-Droid 2.0 - The open-source Android app store has been redesigned, with key parts rewritten in Kotlin Compose, and now runs background updates using Android's newer installer APIs.
Sponsorship opportunities
I'm looking for sponsors for both the podcast and this newsletter. Details of podcast sponsorship can be found here. Feel free to reply to this email to discuss further.
The Podcast 🎙
The first of my two episodes with Gui Ferreira is now out - episode 89, all about Claude Code. It's a deep dive into Claude Code (Gui has two Dometrain courses on it!), and along the way we swap our favourite features and tips, including /goal, dynamic workflows vs /batch, auto mode (and the surprise behaviour change that caught me out!), Ctrl+G and Ctrl+S, voice mode, /btw, rewinding, path-scoped rules, and /simplify.
The second episode, on building personal software and using coding agents for things that aren't coding, will be out next week.
Also, just a reminder that we have a Discord community for the podcast! It would be great to see you there :)
Podcast Discord community
My Dometrain Courses
Below are my Dometrain courses. You can also find all of my courses here.
Please help me share this newsletter 🙏
If you've made it this far into the newsletter - I'm hoping that means you've both enjoyed it and found it useful. If you can help me out and share with your developer friends at work and on social media, that would be amazing!
Again - feel totally free to reach out to me, and let me know your thoughts on the newsletter. And see you back in your inbox next week for the next edition! 👋
My socials...
X/Twitter (I've started using this again more proactively now)
Bluesky
LinkedIn
Discord
Mastodon (rarely used)
Or just go to danclarke.com for links to all my stuff 😊
