|
Hello Full Stackers!
We've got 1,423 new subscribers this week, welcome to the stack! 🥞
Two weeks of news to catch up on. Between this newsletter's upcoming first birthday and the biggest launch I've ever done, I've got a lot to share. Opus 5 launched in the meantime, rounding out the latest wave of model release, so we'll take a look at how the dust is settling. I've also added a fun new section for things you can try this weekend.
Here's the plan:
| • | A totally new way to take my courses (plus Full Stack Mastery is officially live, with a deadline) | | • | The model roundup: every major release, judged from actual use | | • | Things to try this weekend: the best techniques crop in months | | • | The Sizzle Reel: everything else that shipped | | • | Meme of the week 😂 |
Let's do this.
| 🍳 Fresh Off the Griddle (My stuff) |
A completely new way to start my courses
Once you've built your Claude Code setup and connected it to all your tools, you become allergic to clicking around crusty UIs. All my courses still relied on said crusty clicking, so I built a dedicated agent tool that launches any of my courses with one command, right inside whatever AI app you already use.
If you've been meaning to try any of my courses, it's never been easier: paste this into Claude Code, Codex, or Cursor, and the AI will set you up and walk you into any lesson
Fetch fullstackpm.com/cli, install it, and help me get started
This is also how you can get a snazzy ✨certificate✨ you can share straight to LinkedIn.
Full Stack Mastery is officially live
After four months of running Claude Code for PMs: Mastery with 750+ members, I've expanded it to work in Codex and Cursor. This is the official launch of Full Stack Mastery 🎉
Membership includes:
| • | 35 interactive lessons across 8 modules that work in Claude Code, Codex, and Cursor | | • | Every new course and weekly premium as long as you're subscribed (AI Building and AI Agents coming later this year) | | • | An active Slack community with 750+ product people from companies like OpenAI, Airbnb, Meta, and Google | | • | Two live sessions with me every week |
After Tuesday, Aug 11, the price goes from $200/year to $300/year. The button below applies code 200FOREVER automatically, and you're grandfathered in at $200 forever.
⚠️ Try 100% risk free: 14-day refund, no questions asked.
Learn more Many members expense this. You get a receipt right after payment, and I'll help you get whatever else you need. |
Everyone who signs up gets the first paid module of the course free! Session will be recorded and sent to all sign-ups. |
| 📊 The dust settles on model season |
The last wave of model releases is finally over: Claude Opus 5 shipped at half the old Opus price (7/24), Moonshot released the Kimi K3 weights (7/27), and OpenAI cut GPT-5.6 prices by up to 80% (7/30). I've now had real time with these models doing real work: building course infrastructure, shipping a launch, and writing this newsletter. Here's what these models are like to use in the real world.
Claude Opus 5. The reviews from people who live in these tools have been rough. Steve Yegge went back to Fable, Teknium rolled it back after it broke his agent stack, and Dan Shipper called it "a hard model to love." Opus 4.6 was my GOAT, but I've hated 4.7 onward... including this one. Huge walls of text, a personality that argues with you, and it does work you didn't ask for. BUT it actually is a quite competent model when something smarter is steering it and keeping it accountable (more on that in one second).
Two things worth doing today if you use it:
| • | JJ Englert's response-style prompt fixes most of the chattiness. Copy and paste it. | | • | ⚠️ Opus 5 silently falls back to Opus 4.8 when you hit a guardrail, with no warning or popup (Fable at least shows you an indicator). There's a settings flag that turns the silent rerouting off, this post shows exactly where it is. Flip it so you always know which model you're really using. |
Fable 5. An absolutely wonderful model and the only thing telling me Anthropic hasn't completely lost the plot. It's too expensive to be most people's daily driver, so everyone's converging on this: Fable as the orchestrator, cheaper models as the workforce. Opus 5 kind of sucks to work with directly, but it's great when you don't have to read its outputs and keep it accountable yourself. Getting in the habit of coming up with an orchestration plan with your agent before kicking off a big chunk of work is the highest leverage move in AI right now.
(Sonnet 5 also shipped in this wave. It came and went with a big 🤷♂️)
GPT-5.6. Fast, very competent, and somewhat robotic but in a good way. Wednesday's price cut makes it even sweeter: Luna is 80% cheaper, Terra is 20% cheaper, and Codex auto-review dropped to roughly a tenth of its old price. If you're already using these, that's free money. I've also been driving GPT models from Fable with the Codex plugin, and the combo works better than either alone.
These great models + extremely generous usage limits = meteoric rise of Codex (OpenAI's version of Claude Code). (Remember: I have a free Codex course.)
Grok 4.5. People love this model. It's become the go-to cheap workhorse, it landed in GitHub Copilot this week, and by most accounts it's both pleasant to drive and more capable than its price suggests. Cursor is also the absolute best agent environment today, so I expect fantastic things from Cursor + SpaceX in the not-too-distant future.
Missing from this list... Gemini. Google shipped nothing on the model frontier this window. Where is Gemini 3.5 Pro? People are blaming the fact that Google doesn't have a harness anyone actually uses, so they aren't getting data for how people drive AI agents. It explains why Grok got so good once xAI had Cursor's data, and it explains why Anthropic and OpenAI are so willing to subsidize your usage through their subscriptions: your sessions are the moat. (It also makes me think those generous subscriptions aren't going away anytime soon.)
| 🛠️ Things to try this weekend: reads and repos |
There wasn't really another headline story, so here are some actionable things you can try this weekend to help you get a feel for this next generation of model's new meta.
| • | Slim your prompts. Anthropic deleted 80% of Claude Code's system prompt for the Claude 5 generation and lost nothing in evals. The write-up comes from the engineer who did it, and it explains why Opus 5 fights your old setup: mega-prompts tuned for 4.8 actively degrade it. Run claude doctor to see what your own setup should cut. | | • | Roll your own agent memory. Victor Taelin's setup is a plug-and-play prompt where the agent keeps an append-only log and compresses it as it goes, so the context stays one size and older details gradually fade. It's a weekend project that will teach you more about agent memory than any explainer. | | • | Learn how these models learn. Lee Robinson's walkthrough of pretraining, fine-tuning, and RL is the best plain-English version I've seen. If model behavior feels random to you, this 10 minute read will fix that. | | • | Ask your agent for annotated screenshots. When your agent finishes anything visual, have it attach screenshots with the changes marked up. You review with your eyes instead of digging through whatever it changed. | | • | "Models are getting spikier, not more general": the thoughtful bear case on this whole release wave, arguing that RL on coding may be eating general capability. Most jobs aren't getting replaced any time soon, it turns out. | | • | Buzz: Jack Dorsey's company Block launched an open-source workspace where humans and agents work as equal members behind one signed identity. It's early, but worth a look if you think about where team tools are heading. |
🤖 Anthropic
| • | Record a Skill landed in Cowork: do a task once on your screen, and Claude turns it into a repeatable skill. It's the closest thing yet to a record-macro button for your whole computer. Two gotchas: once per 5 hours on Pro, and it burns tokens. | | • | Claude Code got noticeably faster in long sessions. Run claude update to get the fix. | | • | ⚠️ The +50% usage promo runs through Aug 19. Also, the $100 API credit some of you got enables overage billing, so set a spend cap. | | • | The MCP spec got a major update. Servers are getting much easier to deploy, and agents can now show real interfaces. If you've built anything on MCP, read the announcement. |
🔮 OpenAI
| • | Codex CLI now supports Claude Code plugin marketplaces (!). The harnesses are converging on shared infrastructure. This is great news. | | • | /import migrates your whole Cursor or Claude Code setup into Codex in minutes, so running both for a week has never been cheaper. | | • | ⚠️ Codex context was cut from 372k to 272k without a word from OpenAI. The community caught it in a public PR. |
🎯 Cursor
| • | Cursor on iPad shipped, with full PR review and merge plus an agent inbox. I guess people want this? |
🌍 Models and ecosystem
| • | DeepSeek V4 went GA, and it's absurdly cheap: V4-Flash runs $0.14 in / $0.28 out per million tokens (Opus is $5/$25), with a 1M context and MIT open weights. A fresh update today (7/31) tuned Flash specifically for agents, including native Codex support. | | • | Kimi K3's open weights shipped (7/27), the largest open-weight release ever. You can use it today through most API providers at $3 in, $15 out. |
Says it all:
And since I completely forgot Meta has models as well:
New here, or been lurking? All my free courses are on-ramps to building with AI:
Just paste this into whatever tool you want to learn, and the AI will help you get set up:
Fetch fullstackpm.com/cli, install it, and help me get started
And if you're ready for the whole stack, Full Stack Mastery is $200/year until August 11, then it goes to $300. Or, want to learn more about the program first? Join my live session Thursday.
Keep building,
Carl 🥞
|
