Hello Full Stack PMs!
Welcome to the Weekly Stack, serving up the hottest AI developments fresh off the griddle, specifically for PM builders. We've got 1,520 new subscribers this week – welcome to the stack! 🥞
You picked a good time – there's a lot to cover.
The last Weekly Stack went out on November 14th. Since then:
Three frontier models dropped
MCP became an industry standard
Every coding tool got a major update
Image generation crossed into "wait, that's actually useful" territory
There's no way I can give everything the attention it deserves. So here's the plan: I'll cover two big stories properly, then hit you with a rapid-fire list of everything else.
You'll be caught up by the time you finish your pancakes (it’s not just me eating pancakes right now, right?).
Let's do this.
Got here through a link? Join the stack! 🥞
🏆 First Sponsor Alert!
Before we get into it: The Full Stack PM has its first-ever sponsor, and it's something I actually use every day.
Nimbalyst is the best way for PMs to use Claude Code.
The problem – Claude Code runs in the terminal, so your options are:
Use terminal and Obsidian separately (clunky)
Use Claude Code in an IDE terminal (better, but editing markdown still sucks)
Nimbalyst fixes this. You can edit files with formatting directly. It's what Claude Code should feel like for people who work in documents, not code.
And it's completely free! If you’re already using Claude Code, give it a try.
Or if you have some spare time over the holidays, my Claude Code for PMs course can be completed in about 5 hours — and you'll start the new year with legit superpowers. It’s great in Nimbalyst!
🍳 Fresh Off the Griddle
In case you missed it — I launched the first-ever extension to my Claude Code and Cursor for PMs courses.
The topic: using Google's Nano Banana Pro through Claude Code or Cursor.
You can make everything from user journey illustrations to architecture diagrams, in whatever style you want, and it all looks amazing.
This combines the most powerful image generation tech with the most powerful coding agents. The AI can run multiple variations in parallel and build your reference library automatically. Basically it does all the things in the cheatsheet below for you.
And of course, it's 100% free. Try it out!
What should I cover next?
Or another request? Just reply to this email!
🧠 Story 1: A Tale of Three Models
Underpinning everything this month was a huge leap in model capabilities. The frontier models all got smarter, and we’re starting to see real strategic differentiation between labs.
Anthropic's new flagship. The headline feature is the "effort" parameter — you can now explicitly tell Claude how hard to think. Low effort for quick stuff. High effort when you need it to really reason through something.
80.9% on SWE-bench Verified (!)
67% price drop to $5/M input tokens
Best at: Architectural design, multi-file refactors, sustained reasoning sessions
I've been living in Opus 4.5 via Claude Code for the past few weeks. Something has shifted. People are really "feeling the AGI".
Google came in hot. First Gemini to break 1500 Elo on LMArena.
The interesting tech here is "Thought Signatures" – the model can preserve its reasoning state across multiple API calls. So when you're chaining tool calls together, it doesn't lose track of what it was doing. That matters a lot for agents.
1M token context (the largest)
Best at: multimodal work (video, images), anything requiring massive context
OpenAI's response. They split it into three variants:
Instant — fast, personality-forward
Thinking — the workhorse for most tasks
Pro — maximum reasoning, will think for minutes if you let it
New feature: response compaction. The model can compress conversation history into an encrypted summary, so it can handle ultra-long tasks that would normally blow out the context window.
400k context with near-100% accuracy on long-context retrieval
100% on AIME 2025 (math benchmark)
Best at: reliability, generalist knowledge work
🔌 Story 2: MCP Supremacy
This one flew under the radar for most people, but it might be the biggest news of the month.
They created something called the Agentic AI Foundation, co-founded with Block and OpenAI. Google, Microsoft, and AWS are all supporting it.
Quick refresher: MCP (Model Context Protocol) is the standard for how AI models connect to external tools and data. Think of it as USB-C for AI: one connector that works everywhere.
The numbers are huge:
97M+ SDK downloads
10,000+ community-built servers
Integrated into Cursor, Windsurf, Replit, and basically every serious dev tool
Here's the second-order insight worth thinking about:
SaaS products are going to compete on MCP server quality, not just UI.
If your product doesn't have an MCP interface, AI agents can't use it. And increasingly, the AI agent is the user. Products without MCP support risk becoming invisible.
The race for agentic integrations is on.
Here's a real example: Typefully (the tool I use to schedule X and LinkedIn posts) just launched an MCP server. Now Claude Code can interact directly with my posting schedule. I don't even need to open the app.
This is even more reason to start getting comfortable with Claude Code and Cursor. I'm basically living in Nimbalyst now, with Claude Code handling more and more of my actual work.
😂 Meme of the Month
Normally I do memes at the end, but you’re going to need a break before this next part
Product managers in 2026:
🥓 The Sizzle Reel
Everything else that happened! Brace yourself! As a reminder, I only cover news that’s explicitly practical for builders. New models, tool updates, or capabilities, and not things like OpenAI’s new $10 quadrillion valuation or whatever.
So these are things you could immediately play and build with. What a time to be alive.
Models (The Rest)
GPT-5.1-Codex-Max — First model to sustain 24+ hour autonomous coding sessions. Uses "compaction" to work across multiple context windows.
Gemini 3 Flash — Pro-level reasoning at Flash prices. $0.50/1M input. Now the default in the Gemini app.
DeepSeek V3.2 / R1 — Matches GPT-5 performance at 10-30x lower cost. Open weights. This is the value play if cost matters.
Grok 4.1 — 65% reduction in hallucination rate. 2M token context. Agent Tools API includes web browsing and X data access.
Voice APIs
OpenAI Realtime API (Dec 15) — Stealth-dropped "mini" versions of their voice models. Much cheaper. This opens the door for voice features in apps that couldn't afford the full model.
Grok Voice Agent API — $0.05/minute. Cheapest voice API in the market right now.
Coding Tools
The Cursor vs. Windsurf saga continues.
Debug Mode — Instead of guessing fixes, it runs your code, collects logs, and suggests minimal fixes based on what actually happened. This is genuinely new.
Visual Editor — Drag-and-drop UI changes in a preview, and the AI writes the code. Wild for prototyping.
Multi-Agent Judging — Multiple agents evaluate each other's solutions.
BUT: The rollout was rough. Users reported lost work, settings getting wiped, general bugginess. Proceed with caution.
Got acquired by Cognition (the Devin people)
Users say it's better for large codebases — "understood context better"
$15/mo vs Cursor's $20/mo
Also:
Claude Code in Slack — Tag @Claude in a thread, it reads the context, spins up a coding session, and posts the PR link. TechCrunch called it "a bigger deal than it sounds." I’m personally extremely excited to play with this for my personal Slack over the holidays.
App Builders
Bolt.new — V2 added React Native support via Expo, so you can build apps! Browser-based, no local setup. Great for fast prototyping.
Lovable — Added MCP support and deep Supabase integration for backend stuff. Basically "prompt to SaaS."
Replit — Design Mode turns an idea into a live site in ~2 minutes. MCP support is in. Enterprise data connectors for BigQuery, Snowflake, Databricks.
Research & Productivity
NotebookLM — 8x larger context window (up to 1M tokens), 6x memory capacity. Custom personas now support 10,000 characters (was 500). This thing is becoming a power tool.
Google Deep Research — A Gemini-powered research agent that can ingest entire databases and produce structured reports. Plus an Interactions API for developers. It does amazing research but does’t follow format well ime.
n8n 2.0 — Task runners enabled by default. Added MCP Client Node. New Chat hub for LLM interactions. I have a cool collaboration coming up here.
PM-Adjacent
Amplitude + Kraftful — AI Feedback feature ingests support tickets, G2 reviews, Gong calls, surveys — clusters them into actionable themes automatically.
Linear Agents — Auto-triage incoming tickets, deduplicate against existing bugs, file properly formatted issues.
Dovetail — "Magic Clustering" of research notes. Query across hundreds of transcripts with natural language.
Image Gen
Nano Banana Pro — Unbelievably good. If you haven't tried it yet, check out my new Cursor and Claude Code courses to see how you can use it as a PM.
GPT Image 1.5 — OpenAI's native image generation is competitive again. Good for mockups and quick visuals.
PHEW! SO MUCH STUFF!
🥞 The Last Pancake
If you have any time over the holidays and you haven't done either the my Claude Code or Cursor for PMs courses — choose one and do it. Cursor has nicer UX, but I personally find myself using Claude Code more, so that’s my recommendation. Both are great.
About 5 hours total.
You'll start 2026 with unbelievable new abilities.
Your life will never be the same.
I beg you.
For your own good.
Keep building and happy holidays!
Carl