Build big, then protect your keys
Testers who have used Fable 5.2 for a couple of weeks call it a dramatic new step, and it puts almost any kind of software within reach of a layperson. It also costs about twice as much as Opus 5.5, and the same ambition attracts thieves who want your API keys. This session shows what Fable 5.2 is building, how to spread tokens across models, and the security habits that protect them.
Key takeaways
No software is too ambitious for Fable 5.2. Testers describe a dramatic step in capability. It rendered Washington, DC and the Colosseum in pure Three.js code, and people used it to build a working ambient instrument in four days, a wine-making game and a record player whose crackle is synthesized from modeled grooves. It costs roughly twice as much as Opus 5.5, so keep it for the hardest and most creative tasks, use Opus 5.5 to orchestrate and Sonnet 5.5 for easier work.
Thieves want your API keys. A stolen credit card has spending blocks. A stolen Anthropic or Gemini key can spend hundreds of thousands of dollars within seconds, and the tokens get resold at a discount or used for model distillation. Set spend limits, skip auto top-up, never paste a key into chat, keep keys in a gitignored .env file and store them in a password manager such as 1Password, which now offers 1Password for Agents.
Two cheap checks close most of the gap. TruffleHog searches your code for leaked credentials, and a scan of one laptop turned up .env files sitting in a backup. A seven-day minimum package release age on npm, pip and the rest means a new package waits until the community has had time to find and pull a malicious release.
Tools covered
- Claude Fable 5.2: Anthropic's most capable generally available model for the hardest coding, reasoning and creative technical work.
- Claude Opus 5.5: Anthropic's leading model for complex coding, AI agents, and professional work.
- Claude Sonnet 5.5: Anthropic's faster, lower-cost model for well-scoped coding, agents, and knowledge work.
- GPT-6.1 Sol: OpenAI's upgraded Sol model for complex coding, computer use, and professional work at lower cost.
- T3 Code: Open-source interface that runs Claude Code, Codex and other coding agents from one place, on desktop or phone.
- 1Password for Agents: Lets personal AI agents sign in to sites and apps with your approval, without the agent ever seeing your password.
- TruffleHog: Open-source command-line scanner that searches code, files and history for leaked credentials and API keys.
Fable 5.2 and security hygiene, in full
Fable 5.2 is a new step
We'll start with the fun stuff and end with security. There's someone I love to follow for new releases who works at Arena, one of the really legit benchmarking companies. He does 3D tests on models. Every time a model comes out he has early access, and he runs these 3D prompts he's created and burns, I'd guess, $20,000 worth of tokens creating them. I'm not going to burn $20,000 on tokens, so I'd rather watch a YouTube video to see what's new.
The short answer from the people who have been testing Fable 5.2 for a couple of weeks is that it's a dramatic new step in model capability. Ethan Mollick, who you should definitely follow, is one of my favorites on AI applied to teaching. He says it really feels like a jagged AGI moment. It's not great in every way, but it's a fantastic model all around, and certain capabilities feel like a whole new thing.
Here's one of the tests. It's a model of Washington and the White House. It's not that Opus 5.5 isn't great at 3D, but this looks super realistic: the entire city of DC, with rendered trees, all written in Three.js, so it's pure code. I'll jump through a couple of others so you can see the side-by-sides. Here's Fable 5.2 doing the Colosseum, which is also pretty incredible, and Opus 5.5 was the next best behind it. It's not that everything is about 3D, and the models are definitely being trained on 3D now because people like to demo it. It's just a good shorthand for seeing what's possible.
You'll remember from earlier sessions that when Fable came out, it was an amazing model that none of us could afford to use. We saw the progression: unaffordable for a couple of weeks, maybe a month or two, and then they release smaller models that perform at almost that quality. For now it's very expensive, about twice the cost of Opus 5.5, but for certain tasks you'll want to try it. I think those are the hardest tasks, especially ones where you need some creativity, because it is a creative model.
No software is too ambitious
As a designer, I find these legitimately beautiful, poetic worlds. This is the first time I'm seeing something I'd call quite tasteful. He's getting free tokens, so he boiled the ocean: 25,000 agents created over 450 worlds inspired by books, art and places, all in code. You can play with them, and the GitHub is open source. Here's someone going through Leonardo da Vinci's notebooks and building all of the inventions as working 3D prototypes, which is also pretty incredible.
Then there's Kieran Klaassen, an engineer who has also made music his whole life. He built his dream instrument in three or four days, an instrument that creates ambient music and connects to his Ableton sampler, so he can jam with it on the keyboard. It has different instruments, like a reverb cascade. This is full-on music production software built in four days. You can build some of the most complex software out there with Fable 5.2. Right now there isn't software that's too ambitious for these models. Almost any software you've interfaced with, you can start going after as a layperson, which is absolutely wild. I'll keep saying it: challenge your ambition and your dreams, and ask how big you want to go and how crazy you want to be.
Here's one with a learning experience. He's a fan of wine and wondered what a slow educational game about making wine would look like. It creates a beautiful landscape where he can plant his vineyard, with a little spot to ferment wine, and it's all based on the actual science of making wine. Everything is in there, from making bookings to buying supplies to the fermentation process. It's a game built to learn about a subject, rather than reading a book or an article about it. These are new ways of learning and exploring ideas through games and software.
Another crazy project: he built a 3D record player, and then he replicated how a record is encoded, the actual grooves. The audio you hear sounds crackly like vinyl, but it's completely synthetic, modeled on going over something like 100,000 grooves. You're hearing a synthetic replication of vinyl audio. So get weird with it.
Fable 5.2 eats up tokens
In summary, to the people playing with it, this is a game-changing model update, and it opens a whole new category of things to explore. Definitely play around with it, but it will eat up your tokens. In the next couple of weeks I'm expecting an Opus 5.5 that can do things at a comparable level for much cheaper. In the midterm we can come back to building out our orchestrators and routing rules.
I probably wouldn't use Fable 5.2 as an orchestrator. You'd probably still use Opus 5.5 for that, point Fable at a difficult task, a certain feature or some creative front end, and use Opus 5.5 for most of the other work and Sonnet 5.5 for easier things. It's dramatically more expensive.
OpenAI's models, especially with the release of Sol 6.1, are incredibly cheap compared to Anthropic. I think they're feeling some heat, because Anthropic has got its groove back and is putting out incredible models, and I think OpenAI is seeing usage decline. They changed how tokens were calculated, so your $200 a month plan was essentially cut in half, but because the models were so much cheaper you got a lot more for it. So it was a cut, but not a cut. Now they've reverted it back to what it was before. They're dramatically cheaper, but right now not as good. I still use them day to day for writing and for my knowledge work, but for coding and any engineering work it's all Anthropic right now.
The great thing about the Anthropic models is that they're all pretty solid. Even Sonnet 5.5 is very good at design. As a default I use Opus for orchestration, because I need that intelligence to know what's what and to understand the architecture of my project. If I'm working on something net new, I'd have Opus work on it. If it just takes more effort, something to do over time, I'll default to Sonnet and see how that goes, and if I run into issues I can always upgrade. Be careful about upgrading mid-conversation, though: you'll lose your cache and it will cost a lot of tokens. For the most part, Opus for the hard things, and Sonnet does a great job on the easier, more repetitive work. If you're using both ChatGPT and Anthropic models, then similar to Sonnet 5.5, for something long-running where I don't care about speed I'd throw it at Sol 6.1. Sol 6.1 is also great for engineering work.
Put spare tokens to work
On ChatGPT usage limits, which is kind of funny: they often give refreshes. One of the top product managers for Codex will say you get a refresh, and then they just refresh your credits. Midweek you're running out and suddenly you get one, and sometimes you have a couple in the bank. Someone wrote that it's one of the most incredible marketing stunts ever, because they've recategorized surge pricing as a handout. They control cost through price, charging more at surge times, so their default price is probably lower. Then they raise the cost and give you a refresh when it's lower, so you think it's amazing, because surge pricing sucks. This person's view was that it's very well-marketed surge pricing, which I thought was clever. Give them surge pricing and make them love it. It's very difficult to predict, because you don't know when you'll get a refresh.
As for what to do with spare tokens, one thing that's related to what we'll talk about next, which is security, is security audits of your codebases. That can be quite token heavy: reviewing all of your code and looking for vulnerabilities. If you have side projects you wanted to build out, it's great to have a couple on the go and push a bunch of new features. Other than engineering, 3D projects can be very token heavy, and so can motion and video editing. Having projects on the side and throwing tokens at them is a good use of tokens.
I can send a couple of GitHub projects that track your token usage; I'll dig them up. They've removed the five-hour limits as far as I remember, though it always changes, which is why you're only seeing the weekly limit. I think even Claude might have done that for the Ultra plan. There are open-source repositories that calculate your token usage from your chat history. A third option is a different harness where you can plug in all of your subscriptions. T3 Code does that. I haven't used it before, but you can have all of your subscriptions in one harness and flip through them, and probably see your usage too. I'll send some token calculators over.
Thieves want your API keys
I'm going to read this post, because it's a bummer but worth talking about. Claire Vo is an absolute AI gangster and PM extraordinaire, a really cool person who has done really cool stuff and is now a big AI creator in the space. She has a product called ChatPRD, an AI product that helps you write product requirements documents. It's just her and a small team of one or two people. She wrote the other day, in a vulnerable post: "I spend 90% of my time on ChatPRD dealing with increasingly creative international token fraud operations. It makes me want to shut down the company, mostly to steal Opus tokens on resold promo codes, or in the time between the trial start and our fraud alerts infra. This has cost me tens of thousands of dollars, and it gets worse every day. Luckily the company is nice and profitable, but every time it happens it bums me out in extreme ways."
I can imagine the bummer of being hacked nonstop. She goes on that building a product is incredibly fun, that it's the footprint by which she tests all this AI stuff as an engineer, founder and product builder, and that she has a little Grok Bot army that runs operations. They've been pentested multiple times and SOC 2 audited, and work with some of the biggest companies out there. But it's horrible to have a small team spend basically all its time shutting down an Indonesian fraud ring, an Egyptian code farm or Chinese signups on an auto loop for 24 hours. She feels like she's constantly making product calls that trade protecting from abuse against giving users an amazing experience. It has sucked the joy out of her little AI win.
She's also been saying this isn't her first rodeo. She's an experienced product manager and a former engineer, and she does all the safety stuff. But hacking, especially token theft, is a really big thing right now, and it makes having a small AI business really challenging. I bring it up because, as we graduate from vibe coding in our basements to putting real products into the world, especially AI-augmented ones, we're going to have to deal with a lot of fraud.
I'll talk a little about the mechanics. I'm not an expert, but it helps to understand what people want. They're after API keys, especially Anthropic API keys. If I steal a credit card, there are all kinds of security blocks on how much I can spend. But if I steal a good API key from Anthropic or Gemini, or OpenAI too, I can spend hundreds of thousands of dollars within seconds. The money is farmed by reselling those tokens on black markets at discounted prices. I could get an Anthropic subscription for 50% off on some shady website, because for a lot of people the cost of these models is out of reach. Sometimes it gets extra meta, and the stolen tokens are used for model training and distillation. It's a whole underworld. They want your tokens.
Keep your API keys safe
The biggest thing to talk about is how we keep our tokens safe. Maybe you grabbed an OpenAI API key to play around, or a Gemini one to try voice models, or you connected your fal account for generative imagery and video. You want to be very careful with those keys.
First, create spend limits on them, so that if a key is stolen there's a limit to how much money can be pulled from it. Auto top-up is very dangerous and probably not a good idea. I top mine up manually as the money runs out. Second, never paste an API key into chat. The easiest way with an agent is to ask, "Give me a safe way to provide this API key," and it will usually do a decent job. API keys are generally held in a local environment file, the .env file. If you're using Git, make sure that file is gitignored. If it isn't, Git will submit the environment file to your repository, and if you're using GitHub it'll be uploaded with your code.
It has always been a good time to have a password manager, so here's a public service announcement for 1Password. I use it and I'm a huge fan. I've used it for years and I'm not paid for this. Now more than ever a password manager matters, because when you generate an API key you generally only see it that once. It shows you the key, you copy it, and you won't see it again. Then what do you do, put it in your notes app? If you don't have a password manager, I don't know where people would put it, and you'll end up doing bad things with that key. 1Password is a game changer for general password management. I use it for my entire family, so my parents, who aren't very good with passwords, are protected with one account.
Increasingly, up top here, there's 1Password for agents. They're getting with the program, and it's now a password manager for your agents. Grok Bot has the best integration right now, with a 1Password plugin. Every time the agent wants to access an API key, a login to your bank, or your shopping website to make a purchase, it puts up a prompt, you use your fingerprint or whatever, and it gives you access. Given this environment, I really recommend you get a password manager. It doesn't have to be 1Password, there are others out there, but I've used it for a long time and I like it. It's definitely something to think about.
TruffleHog finds leaked keys
The next tool was recommended by the head of security at Cognition, so I trust it. Cognition makes Devin, an amazing coding harness. It's a little too pricey for me, but if you're serious about engineering and need a good software factory, I hear good things. I'll share the article that Jack Cable, their head of security, posted. It's more about security for companies. If you had a small company and wanted to lock it down following best practices, I think it's a really good place to start. He's a credible security person in the space and has good guidelines. One thing he recommended was a tool called TruffleHog. It's a small tool that runs in the background and comes with some skills, and what it does is search your codebase for leaked credentials.
You just install it, with brew install trufflehog in the terminal. Once it's installed it's a CLI, and you ask your coding agent to use TruffleHog to look for leaked credentials. I did it today and thought I was in the clear, and I wasn't. I had a couple of environment files still on my computer, locked up in a backup, and I had no idea. Even if it's not in the cloud, it's not good to have API keys or environment variables anywhere on your computer. TruffleHog is a good way to ask whether a codebase has any of that. If you have anything on the public internet, try TruffleHog to see if there's anything there.
Wait seven days to install
There's another thing, a prompt I put together based on people smarter than me. It's a time lock. I'll read it: "Check which package managers I use on this computer." Package managers are different ways to install software dependencies. They're like software, but for code. In order to build software, you need dependencies, different frameworks that let you do certain things. There's npm, pnpm, Yarn, Bun, pip, uv. If you follow the terminal, you'll have seen some of this happening in the background.
Packages you install through those, like an app store, can often contain malicious code. If you follow the right people on Twitter, every week there's a new big incident. A lot of these vulnerabilities are in tools that are fundamental to software, and you can download malicious software very easily when vibe coding if you just install anything. So this is a prompt you can drop into your agent if you feel so inclined. It puts a time lock on all of these package managers: "Check which package managers I use on this computer, including npm, and set a seven-day minimum package release age."
What usually happens is that a vulnerability is found within a couple of hours, sometimes a couple of days, and by seven days somebody has almost always figured it out and fixed it. So if a package was released less than seven days ago, don't install it until seven days. That's something I've seen a lot of IT security people recommend, and I felt a little guilty about teaching a lot of people vibe coding and not talking about it earlier. Copy this prompt, paste it into your agent and set it up. Sometimes you'll see something hot on Twitter and want to install it, and it will say you can't. Just remember you have the seven-day block and reinstall a couple of days later. You can't always get the latest hotness, but you'll be safe from almost all of these crazy vulnerabilities.
Find what you forgot about
There's one other prompt, which you can run when you have spare tokens, because it's quite involved. It's from Claire Vo, and I have two more like it. Do it at the end of the week when you still have spare tokens. It goes: go through my email, find all the services I've signed up for, including the ones I forgot about, and use browser use or computer use to look through my connected apps and projects for old keys, unused integrations and anything that could expose my data. Check my AI data and old connectors I don't need anymore. Check my repos for vulnerable dependencies. Put it all in a Google Sheet I can comment on, with what's at risk, how urgent it is, and what I should do.
This is a little intrusive, because it goes through a lot of stuff, but it's a good way to find things you've forgotten about that might make you vulnerable. When you have extra tokens, it's something to consider.
To summarize: they want your API keys, so protect them with everything you've got. Use a password manager, I think it's worth it. Set the seven-day block, which will protect you from a lot of vulnerabilities. And one final thing: be really wary of Chrome extensions. There's a lot of shady stuff in the Chrome extension store. Even if you use ad blockers, uBlock is okay according to this guy, but most ad blockers do a decent amount of data extraction. When it's free, you're the one for sale. With Fable 5.2, as well as the other models, you can also reverse engineer a lot of these Chrome extensions without much sophistication, so if you want to recreate one, build your own.
Most models now default to auto mode, and in general you'll want to avoid bypass permissions unless you're feeling really excited. It's really risky. Auto is quite good: most things will be okay, and if there's something especially risky, it should flag it.
The sad reality right now is that if you want to create an AI business, a lot of that effort will go into security and protecting your thing. Hopefully the global footprint of fraud is uneven. A lot of fraud is focused on English-speaking countries like the States and the UK, and some non-English-speaking countries are less affected, I believe, so it may not be as extreme in Brazil or Europe. Still, it's something to be aware of and brace yourself for if you're creating AI products out in the wild. That's all we've got time for today. It's a little drier on the security side, but hopefully some helpful tips. I'm not a professional engineer, but I'm worried about your security, so I'm doing my best.
Lock down your keys and your installs
Protect the API keys and installs on your computer in four short exercises. By the end, new packages wait a week before installing, your keys live in a password manager, your code has been scanned for leaks, and you have a short security to-do list from a credible source.
Tools you'll use: Claude Code or any coding agent, 1Password (or another password manager), TruffleHog
Exercise 1: Set a seven-day wait on new packages
New software packages are a common way for malicious code to reach your machine. Vulnerabilities are usually found within hours or days, and by seven days somebody has almost always fixed them. A minimum release age means your package managers refuse anything newer than a week.
Step 1. Open your coding agent.
Step 2. Paste this prompt:
Check which package managers I use on this computer, including npm, pnpm, Yarn, Bun, pip and uv, and set a seven-day minimum package release age on each one.
Step 3. Ask the agent to show you the setting it changed for each package manager, and confirm each one.
Expected output: a list of the package managers on your computer, each with a seven-day minimum release age.
Stuck? If the agent says a package manager doesn't support a minimum release age, ask it to set up the closest equivalent or to tell you which version to upgrade to.
When you later try to install something released in the last seven days, it will refuse. Wait a few days and install it again.
Exercise 2: Move your API keys into 1Password
When you generate an API key, you usually see it once. A password manager is the safe place to keep it.
Step 1. If you don't have a password manager, set one up. 1Password is the one used in the session, and there are others.
Step 2. Pick one API key you use (OpenAI, Gemini, fal) and save it as an item in your password manager. Delete it from any notes app, chat or file where you pasted it before.
Step 3. Set a spend limit on that key in the provider's dashboard, and turn off auto top-up.
Step 4. Ask your agent:
Give me a safe way to provide this API key.
Follow its instructions. Never paste the key into the chat itself. If it asks you to use a .env file, ask it to confirm that .env is in your .gitignore.
Expected output: the key is stored in your password manager, capped by a spend limit, and kept out of chat and out of Git.
Optional: if you use an agent that supports 1Password for Agents, connect it, so the agent asks for your approval each time it needs a login or key and never sees the password.
Exercise 3: Scan for leaked keys with TruffleHog
TruffleHog searches your code for leaked credentials. Run it on the folders where you keep projects, and on any backups.
Step 1. Install it from the terminal:
brew install trufflehog
Step 2. Ask your coding agent:
Use TruffleHog to look for leaked credentials in this codebase, and in my other project folders and backups. List anything it finds without printing the secret values.
Step 3. For every key it finds, revoke it in the provider's dashboard and create a new one. Then store the new key as in Exercise 2 and delete the old file.
Expected output: a short list of findings, or a clean result. Even a clean result on your code is worth checking against older backups, where .env files tend to hide.
Exercise 4: Read the security guide and pick three actions
Jack Cable, head of security at Cognition, wrote A practical guide to security in November 2026. It's written for companies, and it's a good place to start if you run a small business and want to follow best practices.
Step 1. Read the guide, starting with the actions marked "Immediately".
Step 2. Pick the three that apply to you or your team. Two you've already done in this session: scanning for secrets with TruffleHog and setting a minimum package age.
Step 3. Ask your agent:
Here are three security actions I want to take: [paste them]. Walk me through doing each one on my computer, one at a time.
Expected output: a short written list of three actions, with the first one done.