Last November, something changed in how we build Agentcy. Opus 4.5 arrived, and for the first time an AI model could take a real piece of work in our codebase, read the surrounding files, make the change and check it, without someone walking it through every step. We'd used AI to write code before. This was the first time it felt like handing a task to a colleague.
Eleven months on, that colleague is much better. Opus 5.5 is an extraordinarily capable agent. It works for hours on a problem, keeps the house rules of a large codebase in its head, and tells you plainly when something didn't work.
The builders are good now. The gap was elsewhere.
Different models for different jobs
We don't use one model for everything, and neither should anyone building with AI. Building the software and running it are different jobs, and they want different models.
For building, Opus 5.5 is the one we reach for: the change that touches twenty files and has to respect how the rest of the system behaves.
Inside the product, the work looks very different. Agentcy makes a huge number of small calls to a language model every day, such as summarising an article, tagging it with a sector, pulling the names out of a piece of coverage or drafting a first line for a consultant to rewrite. That's low-level intelligence, and for it GPT 6 Luna still has the edge in our view. It's capable and cheap, and at that volume the price decides whether a feature can exist at all. Running the most expensive model in the world over every article we read would add up fast.
Anthropic's models lead the way on building and planning, but they don't cover everything. The best set-up picks the right model for each job, whichever lab it comes from.
What Opus doesn't do as well is the step before the building starts: deciding what the work is.
Planning is still the weak link
Any developer will tell you the plan is where projects go wrong. Which parts of the system does this change affect? What order should the steps go in? Which piece can run alongside another, and which has to wait? When a task is big enough to split across several agents working in parallel, someone has to break it up, hand out the pieces and check what comes back.
Fable 5.1 is good at this, though its plans still need a review. It misses some of the finer points, especially in UI and UX design and in integrations, where a good plan depends on a deeper understanding of how the systems work and the processes they support. A plan that gets those details slightly wrong produces work that has to be redone.
Fable 5.5 was widely expected by now, and it hasn't arrived. Nobody outside Anthropic knows yet what it will do or when it will land, so what follows is a bet, not a review.
If the next Fable closes that gap, it could become the planner and orchestrator that sits above the rest of the suite: reading the problem, writing the plan, handing the routine pieces to cheaper agents and the hard ones to Opus, and pulling the results together. That would be the missing piece, because the builders are already there. A planner you can trust is what turns a set of capable models into a team. Until one arrives, people still do that job.
The models most people haven't noticed
The coding models get the headlines. Two other kinds of foundation model are still under the radar, and we think they're just as important.
The first is decision models. I wrote about Typesafe's System One models last month. Jev doesn't write text at all. It answers questions with probabilities you can rely on, so when it says 70%, it's right about seven times in ten. For anyone building agents, that's a big deal. An agent that makes a decision based on a chat model's guess is unpredictable; an agent that acts on a calibrated probability can be given clear rules, such as acting on its own above one level and asking a person below it. Reliability and repeatability are what let you put an agent in front of real work, and decision models give you both.
The second is transactional foundation models: AI built specifically for events, such as payments, merchant histories and the long trail of small records a business produces every day. Language models read sentences. These read sequences of transactions and learn what normal looks like, which makes them a natural fit for fraud, credit and anything else where the signal is in the pattern of events.
Where this leaves everyone else
Put those together and you have a builder, a low-cost model doing the everyday thinking inside the product, a decision model and a transactional model. Add a planner good enough to run them and that line-up could change how whole industries work, faster than most people expect, because each model makes the others more useful.
If AI can copy software feature for feature, where's the moat?
An agent team can now rebuild most software from a good description. Show it a competitor's product and within days it will have the screens, the buttons and the workflows running. Features used to take a year to build and another year for a rival to catch up, and that gap has gone.
Features no longer define good software.
Good software is about the knowledge and the processes you don't see. Two products can look identical on the surface and behave completely differently underneath, because one of them knows how the work is really done: which decisions need a person and which don't, what a good result looks like, and the edge cases that only show up after years of doing the job. In our case that's how PR actually works, from which story is worth a consultant's morning to what a journalist will and won't use.
In an AI product, that knowledge lives in the prompts and context you give the models and in the processes wrapped around them. It's what makes the same models produce better work for you than for anyone else. The proprietary data you hold counts too, because it took years to collect and nobody can scrape it.
Everything else is a prompt away from being copied.