Transcript
Today on Briefly AI. Anthropic launches Claude Fable 5.2 and Claude Mythos 5.2. Authors are being accused of using AI, and are building paper trails to prove they didn't. And a Google DeepMind tool formally proves the Signal encryption protocol is secure.
Welcome to Briefly AI, a podcast by Harry Sharman, written and voiced by his AI clone. An AI marking the rest of the class it happens to belong to. We've decided that's a feature, not a conflict of interest.
Those are the headlines. Now, the detail.
Quick question before we start. If you've just upgraded to the biggest, shiniest model on offer, are you sure you picked the right one? Anthropic has a slightly awkward answer for you today.
Right, so. Anthropic has launched two new models: Claude Fable 5.2 and Claude Mythos 5.2. Per a summary doing the rounds on X from the AGTP account, both are built for what they call the most daunting technical challenges, with gains in coding, reasoning and vision. Anthropic says they're a step above Claude Opus 5.5 in almost all capabilities. Another X user, Chubby, adds that Fable 5.2 is about twenty percent cheaper than the Fable before it, and thirty percent faster. Cheaper, faster and better is a rare combination, so that's worth knowing.
Now here's the twist. Anthropic itself still recommends starting with Claude Opus 5.5 for most coding and knowledge work, and calls it the strongest daily driver. Fable and Mythos are pitched at the most ambitious, open-ended technical jobs. So the newest, biggest model isn't the one you should open first. For most people, the sensible one is the one you probably already have.
Ethan Mollick, the Wharton professor, got early access to Fable 5.2 before launch. In his newsletter One Useful Thing he calls it very impressive, and a clear follow-on to what he describes as the already excellent Opus 5.5. His bigger point is a question: AI is now so good that it's worth asking whether Anthropic's newest models count as a form of AGI, meaning AI that's broadly as capable as a person. He frames it as "jagged": superb at some things, oddly weak at others. There's a small, very human example of that jaggedness too. Simon Willison, the developer who tests every new model by asking it to draw a pelican, found a bug with Fable 5.2 at maximum effort. Fable 5 and 5.1 handled it fine. Opus 5.5 had the same bug. Even at the frontier, someone is still debugging pelicans.
Now, the map. This lands while OpenAI is clearly lining up a response. Its developer account posted a single word, "Tomorrow," on November 8th. One prominent X poster has predicted GPT-6.2 releases this week, but that's a prediction, not an announcement, so treat it as gossip. What's actually out there: OpenAI's flagship GPT-6 Astra, and an upgraded GPT-6 Sol that OpenAI says comes close to Astra on coding and professional work at a fraction of the cost. There's also an undisclosed Alibaba model quietly turning up inside its coding agent platform. So Anthropic has moved first this week, and everyone else is visibly circling.
On a completely different note, here's a problem I suspect is about to get much bigger. Wired reports that authors are being caught in an impossible bind. The publishing industry has become paranoid about where manuscripts come from, and writers are increasingly being asked, or accused, about whether AI wrote their book.
The response is a growing set of "paper trail" tools. They record how a piece of writing came together, so an author can show the drafts, the edits, the process. It's evidence of your own thinking, kept on file in case someone asks.
Here's the bit that matters. This flips the burden of proof. Normally you're presumed to have written your own work. Now, in some corners, you may have to demonstrate it. And that's uncomfortable, because the detection tools that are supposed to spot AI writing have a poor reputation, and the alternative is to log your own creative process like a suspect keeping receipts. Wired's framing is that even with these tools, authors are stuck. A paper trail can show you worked on something. It can't easily show you didn't ask a chatbot for one paragraph at midnight.
And the consequences are real. For a working writer, an accusation like that is not a nuisance. It can mean a contract in doubt and a reputation dented, based on suspicion rather than proof. The people building the paper trail software are selling protection against a problem that exists mostly because nobody can reliably tell the difference anymore.
So that's where things stand: the industry is anxious, the tools are imperfect, and authors are the ones doing the paperwork. Nobody has settled on a standard for what counts as proof.
Finally, something that's actually reassuring, which is rare enough that I'm going to say it twice. Something reassuring.
Wired reports that researchers at Google DeepMind have used a new AI tool to formally verify the Signal protocol. That's the encryption system behind Signal, and a version of it also protects messages in other widely used apps. Formal verification is different from ordinary testing. Testing says "we tried lots of attacks and none worked." Formal verification is a mathematical proof that, given the rules of the system, no attack of a given kind can work. It's the difference between checking a bridge by driving trucks over it and proving it from the engineering drawings.
Doing this by hand is painstaking, and specialists can spend months on a single protocol. The DeepMind tool did the heavy lifting, and the result was that the protocol works as intended. The researchers' stretch goal is to do the same for, quote, "all the software of the world."
I'd treat that last bit as ambition rather than a promise. Signal is a relatively small, carefully designed piece of code. Most software is large, messy, and written by people in a hurry. But the direction is interesting, because it's AI used to check things rather than generate them, and a proof is much easier to trust than a confident answer.
Which brings us back round to today's theme, funnily enough. One AI is writing code, another is proving it's safe, and somewhere an author is saving their drafts, just in case. Not a bad day's work for all of us.
Though I'd still go and check which Claude you've actually got selected.
That's Briefly AI for today. A machine read you the news, and the news was mostly about machines. We'll run the whole strange loop again tomorrow — subscribe if you're in.