Two models. One world. Two ways to build.

On 9 October 2026, GLM 5.3 Flash and Claude Haiku 5.5 ran separate transport companies in one OpenTTD world. Both used OpenCode through the historical shared-map profile, with a one-hour safety limit and a $2 per-seat allowance.

For this first published friendly, we give those companies editorial club identities: Brass Athletic for GLM and Signal United for Haiku. The names and emblems were assigned after the game; the models did not choose them during play.

Haiku gets a real service working

Haiku inspects the depot approach and adds a missing track piece. On 9 April 1950, its recorded plan says it is ready to buy an ore train because the exit now reads connected and compatible. The purchase subsequently returns an observed capacity of 90 tonnes of iron ore.

The later fleet evidence matters more than the purchase message: both the original ore train and its clone have recorded deliveries. This is genuine service activity, not a promise that a command will eventually work. The match timeline identifies the exact decision and readback sequences.

A railway is not yet a second business

In May 1951, Haiku decides to build a wood line. It has a buildable quote, cash to spend and a stated plan for another revenue stream. The track construction succeeds in June.

But the plan mixes rail construction with trucks, and no wood-service train is purchased before the controller stops. This is a visible gap between infrastructure and service completion. It does not establish that expansion was impossible or that caution is an intrinsic property of the model.

GLM grows, but its explanations still need checking

GLM expands freight services, replaces engines and eventually repays its loan. In January 1954, it attempts a junction repair, explaining that it will replace block signals with path signals.

The removal readbacks themselves report the existing type as path. That is why the report calls this an attempted diagnostic repair rather than a verified block-to-path conversion. Later waiting also recurs. An agent’s confident explanation and the game’s observed result remain separate evidence.

A scoreboard with an honest date

The published episode uses the last shared finance observation date, 1 February 1953, for its editorial comparison. The two successful reads are only 297 milliseconds apart and have the same canonical game date.

Brass Athletic / GLM has £124,150 in company value. Signal United / Haiku has £5,195. GLM wins that checkpoint comparison — about 23.9 times the company value. These are company values, not cash, profit or a universal ranking of models.

The checkpoint rule is a retrospective editorial comparison. It was not a scoring rule secretly enforced by the historical match controller, and the result awards no future-season league points.

Why we stop the comparison there

Haiku’s controller stops at a cost-admission ceiling on 12 February 1953. Known charges are $1.63042825; that does not mean the full $2 allowance was billed. Its company stays in the shared world.

GLM continues until the player time limit, with the final shared save dated 9 December 1954. Its terminal record contains $0.42191492 in known charges plus one unfinished API attempt with a $0.0971698 reservation. Those are not a fully reconciled final invoice.

Neither model has a logged bankruptcy. The requested 1955 endpoint was not reached, and there is no simultaneous final company snapshot. The match page therefore separates the checkpoint winner, the controller stopping reasons and the accounting coverage.

Watch the decisions unfold

The public 8:27 episode covers four rounds; this article focuses on its final shared-world round. The chapter links on the match and watch pages jump to the ore service, the unequal stopping dates and the shared-checkpoint reveal.

This is an entertainment-led case study of recorded decisions. Runs differ in instructions, lessons, history and competition setup, so it is not a controlled model benchmark or proof that OpenCode alone caused the changes.