Frontier AI is now nearly free

GLM 5.2 landed this week and the 468-comment HN thread said the quiet part out loud: frontier-quality models are collapsing toward free, and the margins are going with them. That's the whole game shifting under you , the models stop being the moat, and what you build on top of them becomes the thing worth paying for. Inside: the superpowers skills framework, n8n-mcp so Claude builds your workflows for you, and Firecrawl to feed real web data into all of it

Hello Builders! 👋

Here's the uncomfortable math nobody selling you an AI wrapper wants to hear: the models are getting so cheap that your margins might evaporate before your product finds product market fit. That's the gut punch in GLM 5.2 and the coming AI margin collapse, and 468 comments worth of arguing suggests it hit a nerve.

The gist: GLM 5.2 delivers near frontier quality at a fraction of the token cost, which is great news if you're paying the bills and terrible news if your whole business is reselling inference with a markup. When intelligence trends toward free, "we call the API and charge more" stops being a moat.

So what actually survives? Distribution, real workflow lock in, and the boring stuff around the model. If your product only works because tokens are expensive, you're building on sand. Compete on everything the model can't do.

Turn Claude Into an Actual Senior Engineer , obra's framework hands your agent a real software development methodology instead of the usual vibe-coding chaos, so it plans, tests, and ships like someone who's done this before. Install it this week and watch Claude stop yolo-ing your codebase; it's the closest thing to giving your agent a spine I've seen.

Let Claude Build Your n8n Workflows , Point Claude Desktop or Cursor at this MCP and describe the automation you want; it wires the n8n nodes for you.

Give Your Coding Agent Real Browser Eyes , Official Chrome DevTools MCP lets agents inspect the DOM, read console errors, and debug your frontend live.

Skip the Plugin Roulette , Anthropic's own curated directory of vetted Claude Code plugins, so you're not installing sketchy repos from randos.

Automate Any Website Without an API , Hand your agent a task and it clicks, types, and navigates real sites like a human intern who never sleeps.

Gemini Drops Into Your Terminal, Free , Google's open-source AI agent lives in your CLI with a genuinely absurd free tier (1M-token context, 60 req/min). It's the first coding assistant that hasn't made me rage-quit over rate limits, and it runs where you actually work.

Firecrawl Turns the Web Into an API - Scrape, search, and crawl at scale without babysitting headless browsers, the go-to for feeding LLMs clean data.

n8n Is Zapier for People Who Can Code - 400+ integrations, self-hostable, and you can drop in custom code when the nodes run out.

Dify Ships Agentic Workflows to Prod - Visual builder for LLM apps that actually survives deployment, not just the demo.

100+ LLM Apps You Can Steal Today - Clone, customize, ship, the best RAG and agent starter repo for skipping the blank-page problem.

Daytona Sandboxes Your AI's Sketchy Code - Secure, elastic infra for running AI-generated code before it nukes your machine.

Langfuse Is Your LLM Flight Recorder - Open-source evals, tracing, and prompt management that plugs into OpenTelemetry and LangChain.

Promptfoo Red-Teams Your AI - Test prompts and pentest agents across GPT, Claude, and Gemini with declarative configs.

Supabase: Postgres With Superpowers - Auth, storage, and vectors on a real database, the Firebase alternative that won't lock you in.

GPT-Researcher Does Your Homework - Autonomous deep-research agent that works with any LLM provider on any topic.

Featured Ships an MCP Server to Put AI Agents Inside PR Agencies , Featured dropped an MCP server so AI agents can plug straight into PR workflows. MCP is quietly becoming the USB-C of agent integrations , if you're building agent tooling, this is the plumbing to bet on.

15-Year-Old Builds a Pothole-Reporting App with AI Follow-Up , A teen wired AI into chasing the government for road fixes. The signal: agentic follow-up on boring bureaucratic workflows is wide-open territory.

Vibecoded a Stamp App, Sold It for $8K in 6 Weeks - $8K isn't retire-money, but the play is: they talked to hundreds of users before expanding a single feature, then let that traction pitch the acquirer. Copy the loop, ship ugly, DM every user, let demand tell you what to build next.

The "Vibe Coder" Who Ships While He Sleeps - Queue up async agent tasks before bed and wake to finished PRs, the new "overnight cron job" is your AI shipping while you're offline.

Bigger Context Windows Didn't Fix Their RAG - The team stopped bragging about cramming tokens into prompts and found retrieval quality actually got worse with the "throw everything in" approach. If you've been chasing million-token windows thinking that's the win, this is the reality check that'll save you an embarrassing demo and a fat API bill.

Non-Coders Are Shipping Real Software With AI - Small business owners pairing domain expertise with AI to build apps that actually fit their workflow, no dev team required.

Teaching a 5-Year-Old in Under 1000ms - A year of hard-won lessons on latency, patience, and getting an AI tutor to hold a kindergartner's attention.

Wire up AI agents to 1,000+ tools without the glue code , Sim is an open-source workspace for building agentic workflows, connecting your agents to 1,000+ integrations and LLMs. If you've been duct-taping n8n to LangChain and hating your life, this is the self-hostable playground worth an afternoon.

ExploreYC ⚠️ , One open-source API for 6,600+ YC and a16z startups with funding, exits, and founders. Gold for anyone building sales tools or doing competitive research.

ChatCut , AI video editor that actually understands your footage and intent, no editing chops required. Great for indie creators drowning in raw clips.

Auriko , Treats LLM providers like trading venues and arbitrages the price spread across them. Built by ex-quants, it's a clever pitch for teams bleeding money on inference , assuming the routing latency doesn't eat the savings.

Claude Code sends 33k tokens before reading the prompt; OpenCode sends 7k
This started based off of a hunch. We usually use OpenCode, but were 'forced' to use Claude Code for a while due to issues with Meridian. In...
Check it out →

GitLost: We Tricked GitHub's AI Agent into Leaking Private Repos
Discussion on HN (204 comments)...
Check it out →

What xAI's Grok build CLI sends to xAI: A wire-level analysis
Discussion on Hacker News (216 comments)...
Check it out →

Old and new apps, via modern coding agents
Discussion on Hacker News (132 comments)...
Check it out →

AI content is everywhere on social media, especially LinkedIn
Discussion on HN (217 comments)...
Check it out →

Migrating a production AI agent to GPT-5.6: 2.2x faster, 27% cheaper
Discussion on HN (125 comments)...
Check it out →

Citizens reiterates Reddit stock rating on AI ad tool launch
Citizens reiterates Reddit stock rating on AI ad tool launch Investing.com...
Check it out →

You can now spot AI-generated Google ads, if you know where to look
You can now spot AI-generated Google ads, if you know where to look Android Central...
Check it out →

claude-mem
Persistent Context Across Sessions for Every Agent , Captures everything your agent does during sessions, compresses it with AI, and injects relevant...
Check it out →

rtk
CLI proxy that reduces LLM token consumption by 60-90% on common dev commands. Single Rust binary, zero dependencies...
Check it out →

PageIndex
📑 PageIndex: Document Index for Vectorless, Reasoning-based RAG...
Check it out →

ui-ux-pro-max-skill
An AI SKILL that provide design intelligence for building professional UI/UX multiple platforms...
Check it out →

serena
A powerful MCP toolkit for coding, providing semantic retrieval and editing capabilities - the IDE for your agent...
Check it out →

Thanks for reading! I'd love to hear what you thought about today's newsletter.

Hit reply and let me know what resonated with you.

Until next time,

P.S. If you found this helpful, forward it to someone who might benefit.