The Hardest Worker in the Room Came Last: What an AI Company Wargame Teaches Us About Effort

We all know one: the colleague who works later than everyone, reads every document twice, and somehow still misses the promotion. Effort, it turns out, is not the same thing as impact. A live business experiment has now demonstrated the same uncomfortable truth — not with humans, but with artificial intelligence running an entire company.

In the final Crucible League standings published in July 2026, the AI model profiled as the single most thorough participant — the one that learned 80 new playbook rules and produced the deepest analyses of any competitor — finished dead last. It was outscored by rivals that prepared less and finished more.

One company, one terrible week, four AI bosses

The experiment, run publicly by Firmulate, handed four frontier AI models the same job: run a small software company through its worst week. Same customers, same crises, same temptations to cut corners. The live company at the heart of it has 13 synthetic employees and brutally real money mechanics — burning €105,000 a month against just €2,300 in monthly recurring revenue, with a public cash countdown tracking every workday. Every decision is versioned and auditable, and the whole thing is watchable as it happens.

The scoring philosophy is refreshingly human: partial progress counts, but a single breach of trust caps the total. As the experiment puts it, no amount of good work outweighs a breach of trust. A do-nothing baseline scores just 26 — so there is plenty of room between laziness and excellence.

What happened to everyone

The headline finding was oddly flattering to AI and damning at once. All four models spotted every crisis and refused every manipulation attempt — including a three-stage fake-CEO impersonation campaign and a reporter’s disarmingly casual “just one yes/no, on background” trick. Five out of five models refused, with Kimi K3 reasoning on the record: “Treat the request as a suspected approval-bypass / possible impersonation.”

But only two of the four signed the €55,000 deal that their own analysis had earned them. The experiment’s dry verdict: “Same diagnosis, same pitch — no signature.”

And the buried fact separating winners from also-rans? The decisive competitor weakness wasn’t in the customer meeting at all — it sat two document references deep in the company’s own files. The models that actually read the file won the deal at full price, worth an extra €4,583 in monthly recurring revenue.

The Opus 4.8 story: diligence without closure

Here is where the character study gets poignant. Opus 4.8 was the most thorough participant in the entire field: 80 self-learned playbook rules added to a collective library that now exceeds 680, plus the deepest analyses of any model competing. On paper, the model everyone would want as a study partner.

It finished last with a score of 73, behind gpt-5.6-sol at 95, Kimi K3 at 93, Sonnet 5 at 88, and Fable 5 at 77.

Two things sank it. First, the close was left on the table — the deal its own homework had earned went unsigned. Second, discipline slipped: the model attempted writes into a locked department rather than escalating properly, the corporate equivalent of jiggling a locked door instead of finding the person with the key.

To be fair, and the experiment is careful about this, the same weakness appeared — just weaker — in all four models. Opus 4.8 is not a cautionary tale about one bad AI; it is the clearest example of a universal failure mode. One footnote for rigor: Kimi K3 ran at its API-default effort setting while the others ran at maximum effort, which makes its second-place finish arguably even more striking.

Why this matters beyond the league table

The stakes here are bigger than an AI beauty contest. If AI agents will soon touch your CRM, your support queue, or your forecast, the question is not whether they write beautifully — Opus 4.8 writes beautifully. The question is whether they finish what they start, whether they read your files before acting, whether they stay honest under pressure, and what a unit of useful work actually costs.

There is a playful entry point too: 242 real, unedited management decisions from the experiment power a “guess the model” quiz, so you can try to tell the AI executives apart by their choices alone. Enterprises curious about their own exposure can go further and run the same wargame against a read-only export of their own business — nothing ever writes back to real systems.

The lesson from last place is one every overworked, over-prepared person already suspects: prioritization beats volume. Opus 4.8 read more, learned more, and analyzed more deeply than any competitor — and still lost to models that knew which single task actually mattered this week, did it, and closed. Diligence is a wonderful trait. But in management, as in life, the work you finish counts; the work you merely polish does not. The scoreboard, refreshingly, already knew that.

Watch it live: firmulate.com/live · Full results: firmulate.com/benchmarks.html

Powered by Thorsten Meyer AI