Hello friends!
I hope you all had a wonderful month? Wow yeah, it’s been a while.
Sorry for slacking on the emails, it’s been a mix of personal issues and a bit of procrastinating by taking advantage of these codex resets whoops…
I’ll show you all a preview of one of my projects soon!
But enough yap, let’s go over what happened.
Everything GTM. One platform.
Small teams don't have time to stitch together five tools and hope it works.
Apollo gives you everything you need to find leads, reach them, and close deals — all in one place:
230M+ verified contacts
AI-powered outreach
Data enrichment
Inbound lead capture
Meeting scheduler
And more
Stop juggling tools and start building pipeline that scales.
With Apollo, the AI revenue engine powering 4M+ users.
Learn AI in 5 minutes a day
You don't have to scroll every AI thread, track every new tool, or watch every demo.
The Rundown AI breaks it all down for you — the latest AI news, tools, and tutorials in one free 5-minute email every morning.
Trusted by 2M+ professionals at Apple, Google, and NASA.
Vercel bought Better Auth
Better Auth is the open source TypeScript auth library basically half of you have installed (4.7M weekly downloads, 850+ contributors) and Vercel bought it.
Does anything change today?
No. Library stays free, MIT licensed, same name, same team.
Why'd Vercel actually buy it?
Expand their ecosystem: Vercel keeps buying the open source tools that define the Next.js ecosystem (Next.js, the AI SDK, Turborepo). Auth was one of the big missing pieces.
Agent identity: agents currently just borrow your login with no way to scope what they can touch. Better Auth's Agent Auth fixes that, and it's about to feed into Vercel’s AI tools.
What to look for:
Ecosystem favoritism toward Next.js over Remix, Astro, SvelteKit, and Express.
Possible enterprise feature gating down the line, VC money wants returns eventually.
Possible vender lock-in or not being able to use all features if you’re not using a vercel product.
Lots of new coding models dropped this week
So every company decided to launch some new models while I was on break…
Here's what released and what most developers actually use each for:
GPT-5.6 Sol ($5/$30 per million tokens)
The best overall for more complex tasks or if you’re rich, a good everyday model.
GPT-5.6 Terra (just cut today from $2.50/$15 to $2/$12)
The everyday default for most people that aren’t rich.
Great for simple to moderate coding tasks, but honestly it’s as smart or smarter than 5.5, so definitely good for advanced tasks if you have a brain.
GPT-5.6 Luna (also just cut ~80%, from $1/$6 down to $0.20/$1.20)
The cheapest tier got a LOT cheaper today.
Good for commit messages, regex, small stuff you're reviewing anyway.
Also don’t assume it’s “dumb” just because it’s cheap. It’s as smart or smarter than gpt 5.4.
Kimi K3 ($3/$15, $0.30 cached) — Moonshot's 2.8T open-weight model,
Great for UI work.
Claude Opus 5 ($5/$25, same as 4.8)
Approaches Fable's capability at half the price, now the default on Claude Max. Ships with an effort dial and mid-task fallbacks.
MCP just went stateless (the biggest protocol change since launch)
The Model Context Protocol shipped its largest revision since launch.
TypeScript and Python each just crossed 1 billion lifetime SDK downloads on their own, with the whole SDK family pulling in close to half a billion downloads a month. This touches a lot of production code.
MCP drops its stateful session model entirely, no more initialize handshake, no more Mcp-Session-Id header.
Servers can now run behind a plain round-robin load balancer instead of needing sticky sessions and a shared Redis store, a genuine pain point for anyone who's hosted an MCP server at scale.
Two other pieces worth knowing about:
MCP Apps: a standard way for servers to ship interactive HTML UIs that render in sandboxed iframes, instead of everyone hand-rolling their own widget system.
Tasks: long-running work gets a real stateless lifecycle, tools/call returns a task handle, the client polls it, instead of the ad hoc patterns everyone was already building.
Deprecated: Roots, Sampling, Logging, and the old HTTP+SSE transport are all on a 12-month countdown. They still work, just stop building anything new on them.
Auth also got hardened to align with real OAuth 2.0/OIDC deployments, so hooking an MCP server into Okta or Entra doesn't require workarounds anymore.
AI CAN get too expensive, and Fable proved it
I already covered the enterprise side of this a while ago (Meta's leaderboard, Uber's blown budget, Microsoft cutting seats). But now it's starting to affect all users, not just companies.
Anthropic's Fable model will basically bankrupt you if you're not careful:
$10/$50 per million tokens, which is A LOT (double Opus 4.8)
Claude itself warns Fable burns your usage limits about 2x faster than Opus
You can tell it was getting too expensive when developers started focusing on strategies to save tokens instead of just switching models:
Advisor-executor workflow: strong model plans, cheap models execute.
Prompt caching: Saves up to 90%.
Batch API: half price if it's not urgent.
Tighter prompts: every word costs money.
Model routers: pick the right model for the task automatically.
Looks like Fable's price was the actual ceiling of what people will pay, and proof that cheaper models are finally "good enough" for most work.
Why it matters: "which model" used to mean "the smartest one." Now it means "smart enough for this." Sounds like we’re reaching the optimization phase of AI models.
The React Compiler, actually explained - Neciu Dan breaks down what the React Compiler actually does under the hood, how the compiled output works, and what to do with all those old useMemo calls now that the compiler will happily out-optimize you.
Reverse Engineering ChatGPT Web: How OpenAI Built for a Billion Users - A genuinely deep technical dive into the chatgpt.com tech stack and how optimized it is.
Rewriting Bun in Rust - Bun's creator details how they ported 535,000+ lines of Zig to Rust using dozens of parallel Claude instances in an 11-day sprint, and why compiler errors turned out to be a better feedback loop than a style guide.
Five studies that are changing how I think about AI in software engineering - Brian Houck rounds up five recent papers on AI and dev productivity.
We Stopped Using RSC on TanStack.com - Tanner Linsley explains why React Server Components, which solved a real performance problem for the TanStack docs site, became more trouble than they were worth.
React Doctor - One npx command scans your entire React codebase and hands back a health score out of 100, built specifically to catch those delicious sloppy patterns AI coding agents love to leave behind
Davit - A native, Electron-free macOS app for managing Apple's built-in container platform.
Canvas UI - Open source HTML-in-canvas and WebGL effects for React, Solid, Vue, Svelte, or vanilla JS, for when you want your UI to do something more interesting than just sit there being a UI.
Codex Security - OpenAI's own CLI and TypeScript SDK for finding, validating, and fixing security vulnerabilities in your codebase, built on the same model that powers Codex. Free, open source, npm install and go.
Question for you this week:
Have you been using AI more for programming?
I feel like I’ve been multitasking more and noticing less issues with AI as long as I’m specific. I want to know what everyone’s experience has been so far with these next wave of models.
Anyways, that’s all from me!
Have a great week, be safe, make good choices, and have fun coding.
If I made a mistake or you have any questions, feel free to comment below or reply to the email!
See you all next week.
What'd you think of today's email?
Want to advertise in Sloth Bytes?
If your company is interested in reaching an audience of developers and programming enthusiasts, you may want to advertise with us here.
