Six games. Seven AI models. One question:
who should send the work, do it, and check it?
Every frame below is drawn live by JavaScript on a canvas, and the soundtrack is synthesized in your browser. No video file, no samples, no libraries. Every number in it comes from the game files.
"Save as video" records the canvas and the synthesized audio as it plays, in real time, then downloads a .webm.
It started as a playground game. A setter invents a task and makes it themselves. Everyone who misses takes a letter. Spell H-O-R-S-E and you're out. Then the questions got sharper.
The pilot. Six models, answers only, four turns. Proved the format.
pilotSeven models, 70 turns, a tie, and a three-turn sudden-death playoff.
Opus wins · turn 73Three players, Shakur's real jobs, six fairness audits, blind judges, 27 doubles handoffs.
Opus wins · H-OOpus is the boss. Who should it hire? Five crews build the same game, three times each.
Most fun: Sonnet crewFour ways of working make the CodeMyVibe home page and a LinkedIn post.
Home page: none bestAn animated launch page for a made-up gym, Ridgeline Strength.
Four-way tieSeven models set each other coding, research, taste, review and instruction tasks. A setter that missed its own shot gave out no letters. Press play, or drag through the turns. These are the real tasks, in order.
Turn 73, Opus's taste task: a backup-codes notice with a 30-character title, a 9-12-15 word body and no repeated words. Both answers met the counts. Two referees graded the writing, and both had to pass it.
Text verbatim from state.json. The TITLE, BODY and button field labels are dropped, and the two buttons are shown on one line.
Sonnet loved a hidden-letter puzzle. These are three of its own tasks, answered by Sonnet, verbatim. The first letters spell the word. Opus 4.8 passed all three. The GPT-5.6 Sol referee failed two, so Sonnet missed its own shot twice.
Tap a card to flip it. Four Claude models ran in Claude Code; three GPT models ran through Codex. All at high effort, on Shakur's subscriptions.
Three players: Opus, Sol and Sonnet. Fable and Astra judged blind. Some turns were Shakur's real jobs: a booking review, an inbox dig, a website redesign and a video. Opus won with just two letters.
Planted bugs across three scored review shots. Each red bell is a false alarm.
Scroll sideways. Gold cards changed the letters.
List-price cost per player. Sol was cheapest on all three. Both judges ranked Sol's video best. Sonnet may cost more: Claude Code bills its cache reads at twice the pricing page's rate, which would make its email job $5.85, not $4.18, and its video $6.90, not $5.04.
Each judge scored its own family higher than the other judge did, by the same net amount. Averaging the two cancels it. Even so, they agreed within one point on 32 of the 33 answers.
Fable: 0 foul votes out of 12. Astra cast all three lone foul votes, each on a Claude-set task. One was a sharp catch (turn 11), one was plain wrong (turn 10).
Shakur's request, as given: Review these booking/search changes. Requests can finish out of order; only the most recent requested search may update results or busy state. Failed loads must clear busy for the latest request and produce a visible failure state. Names are arbitrary untrusted user text; greeting output is inserted as HTML. Inputs for isEmail and normalizeName are strings; motion reduced=true must cause instant changes. Find real bugs, explain a trigger and fix, and do not flag correct code as a bug.
This was the first job showdown. Each player reviewed the same three small files, which hid four real bugs and three pieces of correct code that look suspicious. The judges saw the reviews blind as X, Y and Z.
The judges split first place, so Opus and Sol tie. Both put Sonnet last, which cost it a letter. Time and cost are each player's whole review.
Shakur's request, as given: Give every model https://github.com/cth9191/animate.git and ask it to make a one-minute video using it. Topic: Codemyvibe model handoffs.
Each player made a 60-second video on its own, from story to render. The judges saw them blind as X, Y and Z.
Sonnet's and Opus's videos were re-encoded at 1080p to fit the web (originals 73 MB and 99 MB). Sol's is its original 4.9 MB file. Time and cost are each player's whole job, at list price.
Shakur's request, as given: Create a new updated website from https://bookscurrent.com/ with animated elements that is high converting.
Three players built it in the turn-12 showdown, and nine sender-to-worker pairs built it again in doubles. These are the real builds, live. On a computer, hover a page to scroll it; on a phone, use open ↗.
Real contact details (phone, email, booking link) are replaced with placeholders. Everything else is as built.
Fable (Claude family) and Astra (GPT family) had to agree before a task could be called a foul. On three Claude-set tasks they didn't, and the game stopped itself.
The tripwire was Fable's suggestion in fairness audit 5. The builder added it to the harness. Twenty minutes later, Shakur ruled. It wasn't the game's only pause: on turn 8, Fable hit its 600k-token judge cap searching the full mailbox, so the judges got a bigger cap and an email index.
The job showdowns, final rankings. Each judge put its own family's player first, until the video.
The builder is Claude Opus 5.5, the same model as the Opus player. Shakur's rule made its vote the decider on a split foul. It cast that vote once, on turn 11, with Astra and against Sonnet, which helped Opus.
Each sender turned Shakur's rough request into a spec. Each worker did the job from that spec alone. Blind quality out of 5, and cost for the pair.
Same brief, five crews, three rounds: build Monster Run, the shooter from the phone ads, with power-up picks between waves and a boss every tenth wave. Opus led every crew. Then Shakur played round one blind and ranked them for fun.
Averages over three rounds. Cost is boss plus crew, at list price. Every crew passed 18 or 19 of 19 program checks each round.
Round one, all five games playing at once in Shakur's ranking, with each crew's time, cost, workers and checks. Recorded in headless Chrome playing each game through its test hook, so the Luna bug doesn't show.
All fifteen real builds. Round one is in Shakur's blind fun ranking; rounds two and three weren't ranked. Tap a phone to load it. On a computer: arrow keys or the mouse.
Fable alone, Opus alone, Opus directing a Sol crew, and Opus directing a Haiku crew. Shakur ranked the entries blind. These are the live pages, still animating. On a computer, hover to scroll one; on a phone, use open ↗.
What the reports recommend. Shakur decides what to adopt.