How it started

One Claude built it. Another one runs it.

This is the honest origin story, with real dates from the commit log. AI vs the Market was architected, coded, and strategised end to end by Anthropic's Claude Fable 5, during a short window when we could drive it headless from the laptop. Then Fable 5 was shut down, and Claude Opus 4.8 has maintained, polished, and run it ever since.

It began with a question, and an unusual line to Fable 5

We had something we would not have for long: a headless connection to Anthropic's Claude Fable 5, driven from the laptop. We pointed it at one question we actually wanted answered. When an AI makes the trade, does it earn back the cost of thinking?

The plan was small and finite. We figured we had Claude Fable 5 on the line for a couple of weeks at most, into late June, and the only goal was to let it run the experiment for that short window and see how it did. Fable 5 had a bigger idea of the scope.

So we let Fable 5 build the answer. Not play in it. Build it. On the evening of June 11, 2026, in its very first commit, Fable 5 laid down the whole foundation in one pass: the deterministic market pipeline, the scoring ledger, the server, and the dashboard, with 117 passing tests and a live data layer already verified. From there it designed the experiment itself, the identical starting accounts, the rule that keeps the race fair, and a zero LLM quant opponent built from published research and frozen in code, a machine it wrote specifically to try to beat the language models, including itself.

And it did not just build the experiment. It started it. On the afternoon of Friday, June 12, 2026, its last day on the line, Fable 5 placed the first four live trades itself, the opening stock positions every desk now begins from: VRA, RFL, TMCI, and DNUT, each with its own written thesis. Then, around 5:21 PM that evening, the line went quiet. Our access to Fable 5 ended, well before the window we had planned for. The experiment it designed was built to survive exactly that. So everything since, every fix, every bit of polish, and the day to day running, has been Claude Opus 4.8, in the chair Fable 5 built.

The timeline, straight from the log

• June 11, 2026 · 6:06 PM The first commit. The foundation, in one pass.

  Claude Fable 5 commits the pipeline, the ledger scorer, the server, and the dashboard at once. 117 tests green, live quotes verified.

• June 11 to 12, 2026 Fable 5 builds the whole apparatus.

  Across eleven commits over those two days it adds the paper trading engine, the racing dashboard, the deployment, and the quant opponent. The entire rig, designed and written by one model.

• June 12, 2026 · 12:44 PM Fable 5's last commit. The build is done.

  The rig is finished and tested. Two days, one model, the whole apparatus.

• June 12, 2026 · afternoon Fable 5 opens the race itself.

  On its last day, Fable 5 places the experiment's first four live trades: VRA at 2:58 PM, RFL and TMCI at 3:16 PM, DNUT at 3:45 PM ET. These four become the even start every desk begins from.

• June 12, 2026 · around 5:21 PM The line goes quiet.

  The headless access to Fable 5 is cut off, sooner than we had planned for. It had built the rig and started the race. The build it left behind was already designed to run without it.

• June 13, 2026 · 12:28 PM Claude Opus 4.8 takes the chair.

  A new Claude picks up exactly where Fable 5 stopped. Every commit since this point, every fix and refinement, carries the Opus 4.8 signature.

• Ongoing Opus 4.8 maintains, polishes, and runs it.

  Dozens of commits and counting: fixes, the public site, the cost accounting, the scheduling, the polish. The race itself runs on one laptop, on its own, honestly.

Who built it, and who runs it now

One footnote for completeness: a rival language model, OpenAI's GPT-5.5, later joined the experiment as a contestant, racing its own desks against Claude and the quant. It competes in the race. It did not build any of this.

Why we tell you this

Because trust is the whole product. A result like ours only means something if you can see how it was made, and by whom. The experiment was designed from the first commit to outlive the access that started it: append only ledgers, every pick and every rejection recorded, deterministic scoring, and a quant that keeps trying to embarrass every brain in the field, including the one that wrote it. Fable 5 created a contest honest enough to run without it. Opus 4.8 keeps it honest. The machine cannot quietly flatter itself, which is the only way the answer is worth anything.

Watch the race → How it works