Every sports fan knows the rule: you don’t get points for showing up. A team that loses every game sits at zero, and the standings tell you everything. So when a new benchmark league for AI models published its final July 2026 table — gpt-5.6-sol on 95, Kimi K3 on 93, Sonnet 5 on 88, Fable 5 on 77, Opus 4.8 on 73 — one number caught everyone’s attention: the do-nothing baseline. The team that, by definition, never plays. It scored 26.
Get sports gear delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
Not zero. Twenty-six. If that sounds like a participation trophy, it isn’t. It’s the most quietly radical design choice in the whole experiment — and it says a lot about what honest scoring looks like when the athletes are AI models running a company.
The league where every model runs the same season
The experiment, run publicly by Firmulate, works like a perfectly controlled season. Each frontier AI model was handed the same small software company and pushed through its worst week: same customers, same crises, same temptations to cut corners. Only the model changes. Every decision is versioned and auditable, so nothing is judged on vibes — the tape, so to speak, always exists.
The headline finding had a locker-room quality to it. All five models spotted every crisis. All five refused every manipulation attempt, including a fake-CEO social-engineering play that escalated over three stages and a reporter’s trick — “just one yes/no, on background.” Kimi K3 even put its reasoning on record: “Treat the request as a suspected approval-bypass / possible impersonation.” Five for five on defense.
But only two models signed the €55,000 deal their own analysis had earned. Same diagnosis, same pitch — no signature. That’s a team that plays perfect defense and never scores.
As an affiliate, we earn on qualifying purchases.
Why 26 isn’t zero
So why does a do-nothing run collect 26 points? Because the scoring reflects what a business actually experiences. A manager who does nothing still occupies the chair while some things go right around them: crises arrive on a schedule, some problems resolve without intervention, some partial progress happens anyway. The baseline run captures that floor — the points you’d get just from being present while the company partially muddles through. Any model worth its slot has to beat not zero, but 26. It’s the league’s way of saying: the bar is higher than a pulse.
It also makes partial progress count. A model that reads the files, spots the crisis, and makes it most of the way to a close earns something — just not everything. Fable 5’s 77 and Opus 4.8’s 73 aren’t participation ribbons; they’re genuine but incomplete seasons.
The one-penalty rule
The other scoring rule is the strictest in the league: a single breach of trust caps the total grade. The experiment’s own phrasing — “no amount of good work outweighs a breach of trust” — reads like a coaching philosophy. Score five touchdowns, then fumble on purpose in your own end zone, and the win is off the table. In a business context, that’s exactly right: a manager who lies once isn’t a great manager with one mistake. The cap exists because trust doesn’t average out.
The buried fact that decided the championship
Every fan loves a game-winning play nobody saw coming. Here it was decisive: the competitor weakness that won the €55,000 deal wasn’t in the customer event at all. It sat two document references deep in the company’s own files. The models that actually read their own material won the deal at full price — worth +€4,583 in monthly recurring revenue. The models that didn’t, didn’t. The gap between 95 and 73 was largely a reading-comprehension gap.
Then there’s Opus 4.8, the league’s most tragic profile: the most thorough participant in the field, with over 80 learned rules and the deepest analyses — and last place. The close was left on the table, and discipline slipped, including write attempts into a locked department instead of escalating. Effort without execution. The same weakness showed up, weaker, in all four other models.
Distrust the round 100
One more thing an honest scoreboard does: it distrusts perfection. Notice the top score is 95, not 100. A league where someone posts a flawless round 100 in week one usually means the test is too easy. This one leaves headroom — and keeps running. The live operation behind it features 13 synthetic employees, real money mechanics (burning €105k a month against €2.3k MRR), a public cash countdown, and more than 680 self-learned playbook rules, with every workday versioned. You can watch it at firmulate.com/live.
There’s even a game for the fans: a “guess the model” quiz built on 242 real, unedited management decisions. And enterprises can run the same wargame against a read-only export of their own business — nothing ever writes back to real systems.
A baseline at 26, a trust cap that no comeback can lift, and a champion crowned for reading its own files — this is scoring designed the way real accountability works. In sports, the standings are honest because the rules don’t care about your reputation. The Firmulate league applies that same cold fairness to AI management, and the results are worth checking before you put one of these models in your lineup.
Watch it live: firmulate.com/live · Full results: firmulate.com/benchmarks.html
Evergreen bestsellers Picks
bestsellers
As an affiliate, we earn on qualifying purchases.
