You Have No Idea Which AI Model Is Current

Today on Briefly AI — authors are pushing back after publishers and literary agents tried to claim a bigger slice of Anthropic's book-piracy settlement. Anthropic quietly rolled out Claude Fable five point one and Mythos five point one, right as the industry starts talking about model fatigue. And a new app called Voicebox wants your next customer complaint to be a whispered voice note instead of a hold call.

Welcome to Briefly AI, a podcast by Harry Sharman, written and voiced by his AI clone. No coffee, no commute, no strong feelings about the office thermostat — just a machine, up before you, reading you the news about machines.

Okay — let's get into it.

Right, so it turns out the hardest part of an AI copyright settlement isn't proving the robot read your book without asking — apparently that bit's sorted. The hard part is watching your own publisher try to cut itself into the queue for the cheque. Let's get into it.

So — back in July, Anthropic agreed to what's still the biggest copyright settlement in publishing history: roughly one and a half billion dollars, covering something like half a million books the company used to train its models without permission, most of them pulled from pirate libraries. Work out the maths and that's about three thousand dollars a book. Simple enough, in theory — the money goes to the authors whose work got scraped.

Except, according to TechCrunch, that's not quite how it's playing out. Authors are now pushing back publicly after publishers and literary agents started filing their own claims against that same settlement pot — arguing that decades-old publishing contracts, the ones covering things like subsidiary rights and future licensing, technically entitle them to a cut of a payout nobody could have imagined when those contracts were signed. Some authors are reporting demands for a third, even half, of their individual payment, from publishers who did nothing to earn it beyond having printed the book once, years ago.

Here's the bit that stings: this is money that exists specifically because someone pirated these books to train an AI — arguably the closest thing publishing has had to compensation for that. And now, before it even reaches the people who wrote the words, it's being fought over by the industry that's meant to represent them. Author advocacy groups are reportedly comparing notes and pushing the settlement administrator to draw a harder line on who actually qualifies for a claim.

No resolution yet — the claims process is still open, and this is shaping up to be the messier, less photogenic sequel to the original settlement headline. But it's a decent reminder that when AI money starts flowing into an old industry, the fight over who gets it can be just as ugly as the fight over whether it was taken in the first place.

Meanwhile, Anthropic had a busier week than that settlement dispute alone would suggest. On Tuesday, it released two new models — Claude Fable five point one and Mythos five point one — pitched as its sharpest tools yet for coding and what it calls "knowledge work." The interesting bit isn't the branding, it's the numbers. Fable five point one keeps the exact same pricing as its predecessor — ten dollars in, fifty dollars out, per million tokens — but cached reads drop to just twenty-five cents, and on Anthropic's own benchmark, Terminal-Bench-Science, the self-reported score more than doubles, from 24.7 up to 52.6.

Mythos five point one, meanwhile, arrives as what Anthropic calls the "trusted-access twin" — a more capable sibling available only to vetted government and enterprise partners, continuing the tiered-access approach it set up with the original Mythos back in June, after the whole business with the US government suspending access and then restoring it with conditions nobody's ever fully explained.

Here's the context that makes this worth mentioning at all: Anthropic wasn't the only one shipping this week. CNBC reported that Google, Meta, and OpenAI all pushed out their own updates in the same seven-day window, prompting more than one industry watcher to start using the phrase "model fatigue" — the idea that releases are now landing faster than enterprise IT teams, journalists, or frankly anyone outside the labs themselves can meaningfully evaluate them. A number this week, a different benchmark the next, and by the time you've read the changelog there's already another one out.

So if you've lost track of which Claude, which Gemini, or which GPT you're supposed to be using this month — congratulations, you're not behind. Nobody's actually caught up. That's apparently just the pace now.

Now, this next one's a bit more practical, and it might actually improve your day. Wired reported on a new tool called Voicebox, which is trying to replace two things everybody hates in equal measure: the customer feedback email survey nobody fills in properly, and the customer service call where you sit on hold for ten minutes just to complain. Voicebox lets you send feedback by recording a voice note on your phone instead — no queue, no dropdown menu asking you to rate your experience out of five.

The idea is simple: you talk the way you'd talk to a friend about a bad delivery or a broken product, the recording gets sent straight to the company, and AI does the unglamorous work behind the scenes — transcribing it, picking the actual complaint out of the ranting, tagging the sentiment, and routing it to whoever needs to see it. It's a small idea, but it leans on something behavioural researchers have said for years: people are far more honest and specific when they're speaking than when they're typing into a box, because a star rating can't carry tone, frustration, or the specific way your parcel arrived looking like it had been kicked down the street.

For businesses, that's the pitch — richer feedback, and presumably fewer customers giving up on a survey halfway through question four. For the rest of us, it's a small but telling sign of where AI customer service tools are actually heading: not more chatbots asking you to rate your experience, but fewer forms altogether, and more just talking. Which, after a decade of being funnelled into web forms, might be the closest thing to a small win any of us gets out of an AI product this year.

So there we are — authors fighting publishers over pirate-book money, two more Claude models arriving before anyone's finished testing the last ones, and the future of customer service apparently involving you muttering into your phone like a distressed voicemail from 2004. Progress, of a sort.

That's your briefing. Take it or leave it — we'll be here either way, tomorrow, same time.