i've switched to GPT-6.1 Sol, but it hasn't replaced every model i use. in Cursor, i still plan tasks with Opus 5.5, then hand them to Sonnet 5.5. once the task is well defined, Sonnet does the work well enough for me, at a much lower cost.
being able to make those choices is part of why i keep my own Hermes setup. i can change the provider without giving up the workflows i've built around it.
lately, though, i've started wondering whether that freedom is worth the maintenance. i still use ChatGPT for some things. as OpenAI and xAI improve their own products, keeping my separate setup gets harder to justify.
then i think about what i'd actually have to move. the files are only part of it. there's also everything i've built up about how the work should happen: the instructions, the context, which agent handles what, and where each conversation belongs.
i'd want to keep all of that even if i changed models again tomorrow. that's what makes the decision harder. i'm not particularly attached to a provider, but i am attached to a setup that already knows how i work.
i haven't tried Gemini 4 Argon yet. if you have, i'd like to hear what you used it for and how it held up.
enjoy this edition!
The Headline
the week's biggest story
three frontier launches in three days
Anthropic released Claude Sonnet 5.5 on September 28, OpenAI introduced GPT-6.1 Sol on September 29, and Google announced Gemini 4 Argon on September 30. Sonnet's standard token prices stay unchanged; Anthropic reports up to 30% lower task costs through reduced token use. OpenAI prices Sol at $2 per million input tokens and $10 per million output tokens, with cached input at $0.10. Its claim of near-Astra performance comes from its own evaluations. Google's Argon rollout is forthcoming, not general availability today.
The Build
tools, models, and dev releases
- Jevgrep - ask where a behaviour lives in an unfamiliar codebase; this CLI returns relevant files and source excerpts for your coding agent, with Node.js 22+ and a supported provider key required. (GitHub)
- Worktrunk - create, switch and clean up Git worktrees for parallel agent sessions instead of making several agents edit the same working directory. (GitHub)
- Coucou - keep agent sessions and permission requests visible on macOS, Windows or Linux, without switching back to every terminal. (GitHub)
- fframes - write a video in Rust and SVG and render it on the GPU; the repo includes an agent skill and the source project for its own launch film. (GitHub)
Founder Fuel
funding, acquisitions, and market moves
- Photon raises $4.5 million - the startup builds tools for agents inside messaging apps including iMessage, WhatsApp and Telegram, avoiding a separate app download; its developer and revenue figures remain company-reported. (TechCrunch)
- CScale exits stealth with $145 million - its Series C backs optical interconnects for large AI systems, with NVIDIA and Intel Capital joining as strategic investors; the company is designing for failed network links that should not interrupt compute. (Source)
- Anthropic commits $100 million to implementation talent - Claude Frontier Academy aims to train 10,000 engineers by the end of 2027, with initial cohorts drawn from firms including Accenture, Deloitte and Morgan Stanley. (Source)
Longevity & Health
aging, biotech, and health research
- bladder cancer signals in urine RNA - a Nature Medicine study analysed 683 urine samples and reported 95% sensitivity at 90% specificity for localized bladder cancer; this is research evidence, not a routine screening recommendation. (Study)
- retatrutide's phase 3 results - Lilly reports average weight loss of 20.8% at 80 weeks in the 12 mg group of TRIUMPH-2, involving adults with type 2 diabetes and obesity or overweight; the triple-receptor drug remains investigational. (Results)
- five-year melanoma outcomes - an updated pooled analysis tracks survival after treatment given before surgery, with strong outcomes among patients who achieved a complete pathological response to immune checkpoint inhibitors; pooled follow-up does not make the regimens directly comparable. (Study)
The World
regulation, policy, and geopolitics
- California puts people back into AI employment decisions - Gavin Newsom signed worker-protection laws on September 30, including a ban on employers relying solely on AI for discipline or termination. For workplace AI builders, human review needs to be part of the workflow, not an afterthought. (Source)
- Singapore adds capacity behind the AI chip boom - TOPPAN and Broadcom's AST venture held the opening ceremony for a new chip-substrate facility on September 29. It will make the high-density packaging substrates needed by AI and networking equipment; production is scheduled to start in December. More chips also need more capacity in the less visible parts of the supply chain. (Source)
- EU asks how copyright should adapt to AI - the European Commission opened a consultation covering AI use of copyrighted content, alongside piracy and research copyright. Feedback is due November 3. this is an opportunity for AI providers and rights holders to contribute evidence, not a new law. (Source)
Go Deeper
long reads, podcasts, and documentaries
- a model guide for the GPT-6 family - OpenAI | OpenAI. a practical guide to model choice, reasoning effort, prompts, skills and production workflows, useful alongside this week's price changes. (Read)
- the eternal complement - OpenAI's Intelligence Age | Hemanth Asirvatham and Elliott Mokski. the authors argue that advanced AI's biggest contribution may be the repetitive execution behind ambitious ideas, rather than the ideas alone. (Read)
- Personal Computing 2.0 - Imbue | Josh Albrecht. a manifesto for owning your data and computing environment rather than letting a new generation of agents reproduce old platform lock-in. (Read)
stay curious,
AI BEAVERS