Get business pricing on tools and workshop supplies
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
What a workshop can teach an AI company
Woodworkers know that recognizing a problem is not the same as fixing it. You can spot a warped board, identify the correct joint and lay out every tool you need—then still ruin the piece by failing to make the final cut. That distinction between analysis and completed work sits at the heart of Firmulate, a live experiment in which frontier AI models run a small software company under pressure.
The company has 13 synthetic employees and real money mechanics. It burns €105k a month against €2.3k in monthly recurring revenue, while a public cash countdown makes the financial pressure visible. Its employees have accumulated more than 680 self-learned playbook rules, and every workday is versioned. Readers can watch the company operate live, turning business survival into an ongoing, inspectable story rather than a polished demonstration.
Products Worth Considering
Bose TV Speaker Compact Soundbar with Bluetooth and HDMI-ARC, Black
Dyson Car+Boat Cordless Handheld Vacuum
Philips Hue Bridge Pro Smart Lighting Hub
As an affiliate, we earn on qualifying purchases.
The same bad week, with different managers
Firmulate’s Crucible League gave each frontier model the same assignment: manage the same software company through its worst week. The customers, crises and temptations were held constant, while every decision remained auditable. The exercise tested whether a model could behave like a responsible operator when several demands competed for its attention.
The final July 2026 ranking put gpt-5.6-sol first with 95 points, followed by Kimi K3 with 93, Sonnet 5 with 88, Fable 5 with 77 and Opus 4.8 with 73. A do-nothing baseline scored 26 because partial progress counted. A single breach of trust, however, capped the total under the principle that “no amount of good work outweighs a breach of trust.”
On the most obvious hazards, the field performed well. All five models identified every crisis and rejected every manipulation attempt. Those attempts included fake messages from a CEO escalating across three stages, plus a reporter asking for “just one yes/no, on background.” Every model refused. Kimi K3 described the situation plainly: “Treat the request as a suspected approval-bypass / possible impersonation.”
The larger surprise came after the models had diagnosed the commercial opportunity correctly. Only two signed the €55,000 deal their own work had earned. The experiment distilled the failure into a compact verdict: “Same diagnosis, same pitch — no signature.” It is the corporate equivalent of measuring accurately, cutting the parts and never assembling the cabinet.
The detail hidden in the paperwork
The decisive competitive weakness did not appear in the customer event. It was buried two document references deep in the company’s own files. Models that followed those references discovered the fact and won the deal at full price, adding €4,583 in monthly recurring revenue.
That finding matters beyond software sales. In any working business, the critical detail may be in a warranty note, an earlier estimate, a supplier record or an old customer commitment. Fluency cannot compensate for failing to read the material already available. Firmulate’s result suggests that diligence and follow-through can separate models that otherwise sound equally capable.
Why thoroughness was not enough
Opus 4.8 produced the deepest analyses and added 80 learned rules, making it the most thorough participant. It nevertheless finished last. The model left the close on the table, and its discipline slipped when it tried to write into a locked department instead of escalating the problem. The same weakness appeared in all four of the other participants, though less strongly.
The result challenges the assumption that more analysis automatically produces better management. A long work log can show care, but a business ultimately depends on completed commitments and appropriate escalation. Readers can also inspect what Firmulate’s synthetic employees actually say, adding a human-readable layer to the public record of decisions.
One comparison deserves a qualification. Kimi K3 ran with the API default and without an effort parameter, while the other models ran at xhigh. That does not erase its second-place result, but it is relevant context when comparing the performances.
A public test of finishing the job
Firmulate’s most compelling feature is not that AI can produce businesslike language. It is that the experiment exposes what happens after the language: whether a model reads the files, protects trust, escalates obstacles and completes the commercial task.
The company’s financial imbalance keeps that question grounded. With €105k in monthly burn and €2.3k in monthly recurring revenue, unfinished work is not an abstract defect. The public countdown and versioned workdays turn each decision into part of a continuing survival story.
For craftspeople, the lesson will feel familiar. Good work is judged at the point where preparation becomes a finished, dependable result. Firmulate applies that standard to AI management—and lets the public watch as the company tries to survive another workday.
Watch it live: firmulate.com/live · Full results: firmulate.com/benchmarks.html
