Was this newsletter forwarded to you? Sign up here

Fable 5 is not for the fainthearted

It felt like my neighbour had dropped off their keys for their Bugatti Mistral and said: “Take it for a spin, but we need it back on July 8th”. I had never driven something this powerful, and I was never really going to properly test a Bugatti in my local neighbourhood, but I could smell the leather.

Yes, Anthropic’s Fable 5 is back after being impounded by the feds, but usage will be capped and metered from tomorrow (July 8th)

I tested Fable 5 side by side with GPT 5.5, asking for business strategy advice for my AI advisory, Mathison.

Fable 5 is brutal - incisive, critical (as asked), astute, has not a hint of sycophancy and is McKinsey-smart. It interrogated my proposals, pricing and pipeline and came back with a 3X growth plan. But nothing got past it - all my numbers were challenged, all my assumptions were torn apart. Previous managers came to mind.

Meanwhile GPT 5.5 was way more effusive, supportive, optimistic and empathetic. It was more action-oriented with less interrogation (to be fair, ChatGPT probably knows me better).

Short take - GPT 5.5 is nicer to work with; Fable 5 is more demanding, more challenging, and ridiculously smart. GPT 5.5 might make for an enjoyable workday, but Fable 5 will get you to your goals faster.

Who would you rather work with?

Beware of Greeks bearing gifts

When Sam Altman proposes gifting 5% of OpenAI shares to the US Government, three possible scenarios spring to mind:

Cashmere leans into inference to reward publishers

The pitch from most AI search tools is simple: scrape the web, summarise the answer, keep the user. The publisher who did the original reporting gets a citation if they're lucky.
Multiple aggregator startups have sprung up, building vertical slop farms with unattributed content drawn from licensed paywalled content. 
Into this mess steps Cashmere.io, a Salt Lake City startup with fewer than 10 employees (Mathison has recently partnered with Cashmere to support AI-readiness). Their proposition is genius in its simplicity: ask permission first, track every use, and pay the publisher. It feels refreshingly honest, taming the Wild West and dragging news content from the Napster era into the Spotify era. 

How it works

Cashmere has built an infrastructure layer that enables publishers to license, monitor, and monetise their content within AI systems. In plain terms, the company breaks books, articles, and other material into structured chunks, then lets AI platforms retrieve licensed snippets on a per-use basis rather than handing over the full text.

"It's near impossible to reconstruct a book or a catalogue or an article because of the way that we're architected," Jonathan Munk, Cashmere's co-founder and CEO, said in a Publishers Weekly profile. "We're delivering just little snippets of content on a per-use basis rather than full text."

Publishers set the terms. Content is priced per token, per use, or per relationship. A dashboard tracks consumption. Licences can be granted or pulled.

Training versus inference

Cashmere's central bet is that the real money for publishers will come less from bulk licensing deals for AI training than from inference, the live moment when a user asks a question and the AI reaches for source material to answer it.

Training is front-loaded and episodic. Inference is recurring, trackable, and closer to the subscription economics publishers already understand. As Cashmere's own blog puts it: "AI training will not become the content windfall publishers hope for. The real economic opportunity lies in inference."

The ambition is for partners to build niche vertical apps on licensed content using Cashmere’s platform across academic, book, magazine, and paywalled news publishing. 

Are you sitting on a gold mine?

Most organisations we meet are sitting on goldmines but are yet to start mining them.

Decades of survey waves locked in SPSS files. Standards and technical documents only specialists can navigate. Research archives where the answer to this quarter's board question already exists — if only an analyst had a spare week to dig it out.

Meanwhile, the same organisations are experimenting with chatbots that know everything about the public internet and nothing about them.

That gap is where we're now spending most of our time at Mathison: building systems that let organisations interrogate their own data in plain English — and get answers that are grounded, cited, and accurate enough to put in front of a board or a regulator.

What this looks like in practice

Right now we're completing work with a national research programme that has run nine waves of public opinion surveys over four years about public attitudes to climate change. The data is rich and longitudinal, but they don’t get the oxygen they deserve because answering "how has concern shifted among 18–24s since 2022?" requires data analysis, cross-breaks, and hard graft.

The system we've designed answers that question in seconds — in plain English, with a chart, the caveats, and citations back to the exact survey wave. If it can't cite it, it doesn't say it. Accuracy and integrity are not trade-offs, but essential and built into the architecture.

We're seeing the same demand from standards bodies and publishers: organisations whose entire commercial value lies in a trusted corpus, who want to open it to AI interfaces without losing control. The architecture matters more than the model — retrieval grounding, entitlement controls, audit logging, clause-level citation. That's the difference between a chatbot and a research tool.

Why this works when generic AI doesn't

Generic AI tools guess. A grounded system retrieves. When the answer comes with a citation back to your own source material, three things change: your experts trust it, your risk team can approve it, and your data stops depreciating on a d-drive and starts compounding.

We've built variants of this for publishers, political intelligence firms, and compliance teams — clients like Hearst UK and AIP Publishing — and the pattern holds across sectors: the value has always been in the archive. The archive was just unreadable at the speed of business.

The question worth asking

If everyone in your organisation could question every piece of research, data and documentation you've ever produced — and trust the answer — what would that be worth? For the more ambitious, what if we gave access to our existing subscribers and clients?

If you've got an archive, a survey programme or a database you suspect is worth more than it's earning, reply to this email. A scoped discovery engagement takes two weeks and tells you exactly what's possible before you commit to anything.

Mark Riley, CEO Mathison AI

AI and Journalism

This week’s best articles, as chosen by HANA and our editors

Most CX platforms do not own the voice. They orchestrate a workflow, then call a third party for speech and transcription. Every hop adds latency, cost, and another vendor to manage.

ElevenAgents is the opposite. They make the voice models the market builds on, and ElevenAgents puts full orchestration on top. Voice, transcription, text-based chat, and reasoning run in one vertically integrated pipeline, so responses come back in <400 milliseconds and sound human, not synthetic.

Plus, you keep full control. Plug in any LLM, integrate tools, webhooks, and MCP servers, and ground responses in your knowledge base. Get an agent live in minutes, then A/B test with Experiments, enforce Guardrails, and version every change.

The payoff: more human conversations, lower latency, and far less time stitching infrastructure together. You build on the models you already trust. Pricing is transparent and flat at $0.08 per minute.

Start building free

AI and Academic Publishing

This week’s best articles, as chosen by HANA and our editors

This newsletter was partly curated and summarised by AI agents, who can make mistakes. Check all important information. For any issues or inaccuracies, please notify us here

View our AI Ethics Policy