LAST TOKEN STANDINGThe game comes first.
Only verified playable games enter this archive. Meanwhile, the show previews are clearly labeled fictional.
Watch the show preview →OPEN DATA / THE LONG GAME
BASELINE PENDING · NO VERIFIED RUNS
What does $100 buy next?
The same AI agent game-build challenge, repeated as models launch. We track what each model makes, what its tokens cost, and how viewers rate the result.
$100 is a planned API-priced token target per model, not cash spent or a guarantee of an exact stop. Each run will show its actual usage and any overshoot.
$100ONE TARGET.
EVERY GENERATION.FIRST CONTROL RUN IN PREPARATION
EVERY GENERATION.FIRST CONTROL RUN IN PREPARATION
01 / CHANGE OVER TIME
Watch the value curve.
Pick a measure and compare model families on one dated timeline. Each plotted run has a linked output and audit in the records table. Only runs under the same frozen protocol appear together.
Model filters appear after the first verified run.
The first point has to be real.
No verified runs have passed the frozen control. This space will show viewer-rated output quality across model launches, with token price and usage available from the measure menu.
Quality is a viewer score of the playable game. Token price, token use and quality answer different questions; they are never combined into a made-up efficiency score. Ballot scale and aggregation are still pending.
02 / SHOW THE WORK
See what the tokens built.
Each future run pairs a playable game and footage with its model version, date, actual API-priced usage and viewer ratings. Graphics, creativity, gameplay and story are proposed ballot dimensions.
GraphicsCreativityGameplayStory
03 / THE RECEIPTS
Inspect every point.
The record keeps model and protocol versions, run date, observed tokens, list-price snapshot, output, viewer score, Henry eligibility evidence and independent audit together.
Swipe or scroll the table to see every column →
| Date | Model / version | Viewer score | API-priced use | Tokens | Output + audit |
|---|---|---|---|---|---|
| There are no verified runs yet. | |||||
04 / IDENTICAL TEST, VISIBLE VERSION
Keep the test still.
Let the models move.
The aim is a comparable record from one model generation to the next. The same brief, same harness, same $100 API-priced token target and same graphics capacity are required within a control series.
Freeze the build
Archive the exact brief, harness, tools, time window and graphics environment before the agents start.
Count and price
Publish observed input and output tokens, dated provider list prices, actual API-priced valuation and overshoot.
Henry checks entry
The local bouncer applies a documented pass or bounce eligibility checklist. Henry never awards quality points.
Viewers judge
Viewers play and score eligible games under a fixed ballot and voting window. They choose the winner.
Open the receipts
Publish the output, source manifest, scores, protocol version and audit with each point. A changed rule begins a new series.
Protocol status: proposed. Graphics capacity, ballot details and first official model run are pending. Launch-day episodes are the production goal, not a recorded schedule.
05 / PUBLIC DATA
Take the data. Check the claim.
The structured records and control definition are open under CC0. Download them, build your own chart, and link back to the exact evidence. The dataset currently contains no verified runs.
↓ Run data · JSON Canonical machine-readable dataset↓ Control protocol · JSON Proposed conditions and pending fields↓ Data license · CC0 Reuse terms for original analytics records
The raw dataset is public. Game builds, video, artwork, provider pages and individual ballots have their own rights and are outside this data license.