←Ben’s Bites Waybackbensbites.com/p/half-a-sonnet-makes-me-sing30 Sep 2025sent to 143,908opened 38.0%clicks 10,509unsubs 177← 25 Sep · 2 Oct → · random

The newsletter for the technically curious. Updates, tool reviews, and lay of the land from an exited founder turned investor and forever tinkerer.


Hey folks,

Claude 4.5 Sonnet is out133 - the best model on programming benchmarks. It’s better than Opus 4.1 across the board and better than GPT-5-codex in most cases. But in comparison to other models, it doesn’t give you the full picture here. Sonnet 4.5 is in its own league for agentic coding, computer use, and long-running tasks. It’s got much better vision, and you’ll feel it’s much smarter across the board. And it’s much more aligned (wrt to AI safety). Although I’ve watched livestreams where 4.5 didn’t do so well…

I’m loyal to no model, but if you want some stability with “the best AI you can get for $20/mo”, Sonnet 4.5 could hold that for the next few months. (I’m sure Gemini 3 or OpenAI dev day will soon make me eat my words, haha - if you’re going to dev day, tweet me - I’ll be there!)

Mintlify has a new Agent146 to help you keep your docs up to date with AI. You can share any context (code changes via PRs, Slack threads, links and writing guidelines) and it’ll draft a docs PR with changes.

Before the weekend hit, Meta and OpenAI both released a new content feed in their AI apps. Vibes127 in Meta AI is a feed of AI-generated videos in partnership with Midjourney and Flux. The launch videos and the feed mostly have cute (sometimes absurd) animals dancing, but it’s not hard to imagine where it goes from here.

ChatGPT Pulse224, otoh, is a daily personalised feed of new things that might be interesting to you. It’s proactive (ChatGPT messages you first), curatable (you can tell it what to search for) and limited (ends after a few recommendations every day). It works overnight to search for things based on your memories/recent activity in ChatGPT = compute-intensive = only available in ChatGPT’s Pro plan ($200/mo).

You can now buy things on ChatGPT137. Instant Checkout, a new feature, allows Etsy and Shopify sellers to let people buy their stuff via ChatGPT in exchange for a cut they’ll pay to OpenAI. OpenAI claims it doesn’t affect ChatGPT’s recommendations. OpenAI’s recent “how people use ChatGPT” report classifies only 2.1% of queries are related to purchases (half as many as programming, which is 4.2%).

Gemini updated the 2.5 Flash and 2.5 Flash Lite55 models, primarily making them a lot faster and less hungry for tokens. Browser Use found the new flash model performs at par with o352 on their internal benchmarks (but much faster/cheaper). ps: Gemini also released a new "Gemini Live81” model and a “robotics52” model.

Outresearch the competition in minutes. Catch every critical detail, move faster on deals, and delegate complex tasks with Brightwave’s AI research agents. Get unlimited access free for 14 days — faster insights, instant memos, and the edge you need to win. Start your free trial today.59*

*sponsored

🌐 What I’m consuming


⚙️ Tools and demos

  • Scout Monitoring’s MCP174 - Plain-language monitoring. Ask questions like “why is latency spiking?” and get answers right in your coding agent.*

  • Unified Copilot in Zapier104 - A single agent to create any workflow with access to the full toolkit of Zapier.

  • Lovable Cloud & AI80 - Lovable now comes with backend support (powered by Supabase) and special attention to adding AI features inside your app.

  • Excel’s Agent Mode215 - Microsoft now lets Copilot work autonomously in Excel, and it’s better than you’d expect. (how they built it119).

  • Tembo110 - Background agents that plan, code, and review.

  • Integrity130 - Bring notes, canvases and AI chats into one connected workspace.

  • Cell163 - The fastest way for software teams to go AI‑native. (read more102)

Image163
*sponsored

🥣 Dev dish


🌌 On the frontier

  • Cloudflare is launching a stablecoin for agentic commerce, calling it “Net Dollar109”

  • DeepSeek made an experimental version52 of their base model. It’s 50% cheaper for users and 3x to 10x cheaper34 to serve for inference.

  • GDPval64 - measuring AI on real-world, economically valuable tasks. Opus 4.1 is the best model, just a few points away from human-like performance.

via OpenAI GDPval64

🍦 Afters


That’s it for today. Feel free to comment and share your thoughts. 👋

📷 thumbnail creds: @keshavatearth30