by Bryan Chua
I do my real thinking on weekends. Not in meetings — in conversations with an AI. Code reviews, product roadmaps, the architecture calls I don’t have headspace for during the week. For the past year that meant Opus. Claude Opus 4.8, ultracode mode, was my default thinking partner, and it was good enough that I’d stopped asking whether something better existed.
Then I tried Fable 5.
Last weekends I ran it through planning sessions for projects at GoPomelo and for Routero AI, the AI infrastructure product we’re building at Digital China. What struck me wasn’t that it was faster, or more articulate. It was that it caught things — problems I hadn’t even articulated as problems. On one project review, Fable 5 surfaced 19 distinct development gaps in a single pass: structural, architectural, process-level. I’d run the same context through Opus more than once and never gotten close. Not even half. Some of those gaps had been sitting in the backlog as a vague discomfort I couldn’t name. Fable 5 named them.
That’s not a faster tool. That’s a different category of capability.
Claude Is Now Building Claude
Anthropic published something this month I keep coming back to. As of May 2026, more than 80% of the code merged into Anthropic’s own codebase was authored by Claude — up from the low single digits before Claude Code launched in February 2025. Not assisted by Claude — authored. On the most open-ended engineering tasks, Claude’s success rate hit 76% in May 2026, up fifty percentage points in six months.
The research numbers are sharper still. In code optimization, Claude’s Mythos Preview reached roughly a 52× speedup by April 2026, against about 3× for Opus 4 a year earlier — on a task where a skilled human researcher needs four to eight hours to reach 4×. And in one AI safety experiment, Claude-powered agents recovered 97% of the performance gap between weak and strong model supervision, where human researchers recovered roughly 23% over a week.
This is what recursive self-improvement actually looks like from the inside. Not a science-fiction moment — a quiet compounding. Claude is writing the code, running the experiments, and optimizing the models that will train the next Claude. Fable 5 is, in a real sense, a product of its predecessor.
And it’s already doing the thing Anthropic keeps describing: compressing months of engineering into days. At the Fable 5 launch, Stripe reported pointing the model at a roughly 50-million-line Ruby codebase and running a migration across the whole thing in a single day — work it estimated would have taken a team about two months by hand.
That’s not automation. That’s leverage at a scale we don’t yet have a mental model for.
Why This Puts Anthropic Ahead
The question I get most: isn’t OpenAI doing the same thing? What about the open-source models — Llama, Qwen, DeepSeek, GLM 5.2? Aren’t they closing the gap?
After a weekend with Fable 5, my honest answer is no — not on the dimension that matters most.
The open-source models are mildly better than each other. They compete on benchmarks, inference cost, accessibility. That’s real, and for a lot of use cases it’s enough. But the compound-intelligence gap — the ability to reason about what you don’t know, to surface the thing that was never in the prompt — I’m only seeing that at the frontier. And right now the frontier is Anthropic.
The sharpest line in Anthropic’s own research is the gap still left to close: execution versus judgment. Claude is excellent at doing what you tell it. It’s becoming genuinely good at telling you what should be done. The 19 gaps it found weren’t in my brief — it inferred them from the architecture, the patterns, and what was conspicuously absent. That’s judgment, not execution.
The open-source models I’ve tested are still mostly in execution mode. They follow instructions well. They can’t yet tell you which of your instructions are wrong.
So What’s the Moat?
Here’s what I’ve been sitting with. If models keep improving this fast, and the gap for ordinary tasks keeps narrowing, what does competitive advantage in AI even mean?
My answer, right now: the moat is being a better operator.
Not building the model. Not fine-tuning it. Not even proprietary data, necessarily. The moat is knowing how to direct an AI — how to structure the problem, how to design the workflow so the model’s judgment is amplified instead of boxed in. The winners of the next three years won’t be the ones who trained the best model; Anthropic, OpenAI, and a handful of labs will do that. The winners will be the ones who learned to work with a model that’s closing in on research-grade judgment.
Sit with what 80% Claude-authored code really means. It doesn’t make Anthropic’s engineers obsolete. It means the ones who remain operate at a different level — setting direction, validating outputs, catching the edge cases that compound into risk. Operators, not casualties.
AI handles the audit. You handle the strategy. AI handles execution. You handle the judgment about what’s worth executing.
When Fable 5 handed me those 19 gaps, I didn’t just take the list. I interrogated it. I brought context no model has — our team, our deadlines, our partner constraints — and made the calls: which gaps were critical, which were acceptable risk, which were quietly features in disguise. The model gave me a better map. I still had to decide where to go.
The better you get at that collaboration — prompting, structuring, validating, directing — the more leverage you pull out of every model release. And that skill compounds in your favor whether the next model is a 10% improvement or a 10×.
What Changes From Here
Anthropic is honest about where this leads. It lays out three plausible futures: 1) the trend stalls but today’s capabilities diffuse widely; 2) the labs keep seeing compounding efficiency gains, where humans set direction and; 3) AI automates execution; or AI systems themselves become capable of full recursive self-improvement.
I think we’ve just entered the middle one. Fable 5 is the first model that makes the efficiency explosion feel real to me rather than theoretical. The open question is how fast the third arrives — and whether our institutions, our companies, and our own mental models can keep pace.
What I’m doing about it: building Routero AI so that when the models get better, the scaffolding around them — governance, cost control, routing, memory — gets better too. The platform layer isn’t the moat either. But it’s infrastructure for moats. The real work is making sure the humans using these systems improve faster than the systems improve on their own.
Because the day the models close the judgment gap — the day Claude sets direction as well as it executes — the operator advantage matters more, not less. You’ll want to spend the time getting very good at working beside something that’s about to become very powerful.
Last weekend with Fable 5 was a small reminder of how fast that era is coming.
Bryan Chua is CTO of GoPomelo, Filigrain & Digital China and co-builder of Routero AI, an enterprise AI infrastructure platform.
Sources: Anthropic, When AI builds itself · Anthropic, Claude Fable 5 and Claude Mythos 5 · VentureBeat, Anthropic says 80% of its new production code is now authored by Claude