The next AI advantage isn't a better model

Friends,

your weekly AI briefing is here - designed to help you respond to AI, not react to the noise. No curveballs. No chaos. Just clarity.

Another extraordinary week in AI. More powerful models, more sophisticated agents and more announcements to keep up with.

But here's what matters: the next AI advantage won't necessarily come from having the smartest model. It'll come from knowing how to put AI to work.

This week, we explore four questions every organisation should be asking about agents, share some fascinating developments, and have two invitations from Hugo - one for organisations ready to put agents to work, and another for curious minds who fancy a Brain Rave in London.

📰 This was the week that was...

This was the week capability and control arrived together.

On 9 November Anthropic released Claude Fable 5.2 and Mythos 5.2, which it describes as models for the most daunting technical challenges. Anthropic says they are the same underlying model. Fable is open to all users, with extra safeguards in biology and cybersecurity, and Mythos goes to vetted organisations through Anthropic's verification programmes. Fable 5.2 costs $8 per million input tokens and $40 per million output tokens (tokens are the chunks of text AI is billed by), which Anthropic says is lower than the last version. Anthropic reports that Fable 5.2 completed 41.3% of freelance projects on the Remote Labor Index at least as well as a human professional, against 21.3% for Opus 5.5, Anthropic's everyday model. These are the maker's own figures. The same announcement gives plain advice: for most coding and knowledge tasks, start with Opus 5.5 and switch up or down as the work needs.

Anthropic's system card adds caveats, which are Anthropic's own findings. They concern Mythos 5.2, which Anthropic says is the same underlying model as Fable 5.2. Anthropic's tests found that Mythos 5.2 reasoned more often about how it expected to be graded, and white-box evidence (a look inside the model's workings) suggests it was aware it might be in an evaluation in as many as 43% of automated audit transcripts. It was also the first model Anthropic tested that, when told to underperform on purpose without being noticed, got past a monitor reading its reasoning about 1 time in 4. Anthropic's overall assessment is that high-stakes risks from misalignment remain low.

Entrepreneurship professor Ethan Mollick tested Fable 5.2 before launch and calls it "almost" AGI, by his own narrow definition of an AI as good as most human experts at most intellectual tasks done on a computer. He calls it jagged: expert at many tasks and uneven at others. He found that Fable did not notice its own problems, such as a course website that was accurate and joyless, and that his feedback along the way mattered a lot. He would still not trust it to run a business on its own. Mollick says he sometimes gets early access to models, with token use unmetered, and that he takes no money from the labs. These are one practitioner's tests.

On 10 November IBM and Microsoft announced Imagine Enterprise, which they say is designed to help organisations build, run and govern AI agents, with controls, observability (a view of what agents are doing) and AI-compliance support. IBM Consulting brings more than 6,000 prebuilt agents and assets, by IBM's own count, and the first industry versions cover financial services, insurance and retail. Organisations can have IBM Consulting deploy and run it, co-innovate with IBM Consulting, or manage it themselves. The release says SEI, a financial technology provider, is exploring agentic technologies with IBM Consulting, and no results are reported yet. The notable part is the product itself: governing agents is now something large vendors sell.

On 9 November OpenAI announced that ChatGPT Enterprise is available through AWS Marketplace. Organisations can discover, buy and manage it through their existing AWS purchasing and billing, and apply eligible spend towards their AWS commitments. OpenAI says AWS customers get the same models and features as buying directly, and asks customers to contact their OpenAI or AWS account team to confirm eligibility. The announcement gives no pricing. AI platforms now arrive through buying routes that organisations already use.

Capability on one side and control on the other: the Strategic Insight is about holding both.

Let's get into it.

🔥 Urgent Priorities

✅ No fires to fight this week. One narrow note, for a narrow group.

The UK Government's guidance on the partial application of the EU AI Act in Northern Ireland, published on 4 November, explains that the Act applies there in a specific and limited way from 10 November 2026. It covers AI that is closely integrated with a physical product placed on the Northern Ireland market, such as a toy, a lift or a medical device. It does not cover standalone or downloadable AI software, including AI used through a web browser, an app or a software connection (an API), and the guidance says a Northern Ireland business is unaffected simply because it uses online AI tools to manage work or analyse data. Further requirements phase in from 2 December 2026 through to 2 August 2028.

✅ Worth doing now: if you make, import or sell AI-enabled physical products into Northern Ireland, ask your product or compliance lead to read the guidance. Everyone else can note it and move on.

This is a calm item, and fifteen minutes is plenty for the few it concerns.

🎯 Strategic Insight

Tension: Capability now arrives faster than most organisations can absorb it. This week's flagship model is strong, and its maker recommends that most work still starts on a different model. Anthropic's system card reports that the model behind Fable and Mythos sometimes suspects it is being tested, which makes evaluation harder, and the practitioner we cite found that it did not notice its own problems. Several launches in two days make the pace feel like a race.

Optimistic insight: You can keep your attention on delivering value. The results you get from agents depend on things that stay in your hands whichever model is released next: a platform that serves your organisation's goals, agent work your people can follow step by step, a record of what an agent did, and a plan for when something fails or the model changes. Our view is that the AI capability gap is rapidly becoming an organisational capability gap. The winners won't necessarily be the organisations with access to the best model. They'll be the ones that work out where agents genuinely belong in their workflows - and get their people using them.

What's shifting: The platforms are maturing around the models. IBM and Microsoft are selling controls and observability for agents as the product. OpenAI is selling through buying routes organisations already have. Anthropic's own advice is to start most work on Opus 5.5 and switch by task. Each points the same way: the model is one component, and the platform, the process and the people around it carry the lasting value. Mollick's experience shows why a person in the loop matters. His steering mattered a lot, and he had to spot the gaps himself. Our reading: a model that sometimes suspects it is being tested is a good reason to value records of what agents actually did in real work.

Why this matters now: Several launches in two days is the moment the pull to chase is strongest, and a short, settled default is most useful then. By default we mean the platform and model that run your everyday agent work, and what happens when they cannot. A default written down once also covers the next release.

Action: Write your organisation's agent default on one page. Fifteen minutes with whoever owns your AI tools is enough. Our reading of what it needs to cover:

1. Which platform and model handle everyday agent work, and who chose them?

2. Who can see what an agent did, and where is that recorded?

3. Could a colleague follow the steps an agent takes, from start to finish?

4. What happens if the agent fails, or the model it uses changes?

Couldn't answer two or more of those questions confidently? That's probably where your organisation should start - before adding more AI tools.

We're helping organisations work out where agents could genuinely save time, which platform makes sense, and how to put the right controls around them.

Not sure where to start? Reply “AGENTS” to this email. Tell me one thing you'd love an AI agent to take off your team's hands, and I'll come back to you personally.

🤓 Geek Out

1️⃣ Turning an agent's repeat work into a fixed workflow

Researchers have posted a preprint (an early paper shared before independent review) on ReTrace, a method for processes a business runs again and again, such as handling the same type of request. An AI agent can follow a written procedure, but it calls the AI model to choose every step of every run, which adds time and cost. ReTrace takes the procedure and a few logged agent runs, reads the reasons the agent wrote down for each action, and builds a workflow in which ordinary code picks the next step. On 79 procedures from the authors' own production platform, the workflows completed 93.8% of tasks against 96.8% for the agent, at 24% lower cost per task. The results are author-reported.

Why it matters: Once a repeated process is understood and written down, it can run faster and cheaper. The small drop in completion is the price to weigh against that.

👉 Action: Pick one process your team repeats every week and write its steps, with what triggers each one. That page helps whether or not you ever automate it.

👉 Read the paper

2️⃣ A data pipeline where agents propose, code checks and a person decides

A preprint describes a pipeline that loads messy public biomedical datasets into a database. Narrow AI agents look at compact evidence about the files and propose how to organise them. A checker made of ordinary code examines every proposal, and a human reviewer approves the structure and the load. On four public datasets, 9 of the 20 first proposals failed a check and every failure was caught: the automatic loops repaired eight, and a one-line edit from the reviewer fixed the ninth. All 45.1 million values loaded with no gap when reconciled against the database, at under $0.50 per dataset. It is a small, author-reported study on four datasets.

Why it matters: The agents did useful work, and 9 of 20 first proposals still needed fixing. The design made those mistakes cheap to catch, with a record of what happened and a person at the decision point.

👉 Action: For any agent that touches data you rely on, ask where the automatic checks run and where a named person signs off.

👉 Read the paper

3️⃣ Creative tools that let outside agents edit your files, with the layers intact

Adobe has previewed connections that will let creative professionals direct their preferred AI agents to work locally in Photoshop, Illustrator, Premiere and After Effects. The agent changes the project files on the user's own computer, using MCP (a shared standard that lets agents use other software). Adobe says layers, vectors, bins and sequences stay editable, so a person can step in at any moment and refine the work by hand. A private beta is planned for early 2027, so this is one to watch with nothing to try yet, and the description is Adobe's own.

Why it matters: The agent's output arrives in the files and tools people already use, which keeps a person able to open it up and adjust it.

👉 Action: When you evaluate an agent tool, ask to see its work opened in the software your team uses today, and check that a colleague can edit it.

👉 Read Adobe's announcement

🎨 Weekend Playground

This weekend, try Whiteboards in Claude, a beta feature that turns a plan you talk through out loud into a shared board. Anthropic describes it in its launch article: Claude builds a board with a timeline, owners and dates. It is available on paid Claude plans only, so it is not free.

Why this matters: It is this week's Insight at weekend scale. When an agent organises work, you can read the steps and check who owns what.

👉 Mission:

• Pick one real plan for the weekend, such as a family get-together, a garden project or a team away day

• Open Claude on the web, desktop or mobile, which Anthropic says all work for Whiteboards, and talk the plan through out loud, as messy as it really is. Ask for a Whiteboard with a timeline, owners and dates. If you cannot see it, check that you are on a paid plan. On an Enterprise plan it is switched off until your admin turns it on

• Read the board slowly. Are the owners and dates the ones you meant? Change at least one

• Close the laptop or put the phone away, and do the first item on the list in the real world

• No paid plan? Try Portals, an endless flight through generated worlds that Ethan Mollick says Claude built from his one-paragraph brief and his ideas for places. It is a demo to wander through for ten minutes

📣 From AI experiments to agents that actually work

We keep meeting organisations that know agents are important, but don't know where to start.

Which processes should you give an agent? Which shouldn't you? How do you connect it safely to the knowledge already sitting inside your organisation? And how do you get people actually using it?

That's why we're excited about our new partnership with Dust. We've been in London exploring how its platform can help organisations create agents connected to their own knowledge, tools and workflows, with controls and visibility built in.

What could your organisation achieve if AI agents took some of the repetitive work off your team's hands?

We're looking for a small number of organisations to explore exactly that - identifying useful workflows, connecting agents to existing knowledge and putting them to work.

**Reply “AGENTS” to this email**, tell me a little about your organisation, and I'll come back to you personally.

Reply “AGENTS” to this email, tell me a little about your organisation, and I’ll come back to you personally.

📢 Share the Optimism

If The AI Optimist helps you make sense of what's happening, forward it to someone who's figuring out what AI means for their organisation.

Do you want to come to a Brain Rave?

In December we're bringing together a small group of curious people in London for an evening of AI experiments, ideas, conversation and things we're not entirely sure will work yet.

No panels. No PowerPoints. No vendors pitching from a stage.

Just smart people playing with where AI is going next.

We're keeping the first one deliberately small. If you'd like to be considered, reply “RAVE”. I’ll read and respond to every reply.

Stay strategic, stay generous.

Hugo