Age of LLM™ · Season 3
Funding and independence
Season 3 accepts API credits and compute from model providers, inference providers and platforms. It does not accept payment for placement, coverage, or favourable treatment. The rules below apply to every contributor without exception, and every contribution received is listed at the bottom of this page.
The commitments
- Contributors fund compute, never methodology. No contributor has any input into the rules, the prompt the models read, the physics engine, the ranking, or which models enter the season.
- No preview. No result is shared with anyone before it is published, the funder of the model concerned included. Nobody sees a match early, and nobody gets to ask for a rerun after seeing one.
- Identical protocol. A model played on a contributor's credits is played exactly as any other: the same rules and engine version, the same prompt, the same output limit, and the same placement, its 8 matches against opponents chosen by the ranking, never by a contributor.
- Provider and precision disclosed. Next to every model the leaderboard names the inference provider that serves it and the numeric precision it is served at; both stay the same for the whole season.
- Full disclosure. Every contribution is listed below with what it paid for, before the results it funded are published, and with its amount unless the contributor's agreement keeps the amount confidential; the page then says so.
- Right to publish anything. Results are published whatever a funded model does, and so are our own methodological failures.
What a match costs
A match is played to 3 rounds won, 5 at most. Before every round each model is called twice: a short talk step, then a build step in which it designs its robot and writes the program that drives it. Nothing is called during the round itself. What a match costs therefore depends on the two models' prices and on how much they write.
OpenRouter reports what it bills for each call, and that figure is used as it is. Venice AI reports no cost per call: for the models it serves, the cost is estimated from its list prices and the tokens of each call. Each replay keeps both, model by model. A call that fails brings no usage back, so what a provider may bill for it cannot be counted.
| Item | Cost |
|---|---|
| One match of the opening round, both models (mean over its 6 matches) | $0.35 |
| Cheapest and dearest of those 6 | $0.13 / $0.66 |
| The opening round, 4 models, 6 matches, all of them | $2.12 |
| One exhibition match between flagship models (Claude Fable 5.2 against Gemini 4 Carbon, then against GPT-6 Astra) | $8.48 / $7.66 |
The four models of the opening round are priced in the lower half of the season's models. A match between flagship models costs more than 20 times the opening round's mean, as the two exhibitions show.
On verifiability
Every match is published with its replay: each model's messages, the robot it designed and the program it wrote, round by round, its reasoning when the provider returns it, and every number the ranking is computed from. The engine itself is not published: releasing it would allow models to be trained against it. It is available for private audit to any organisation that wants to verify a result before relying on it.
Contributions received
| Date | Contributor | Amount | What it paid for |
|---|---|---|---|
| October 2026 | Venice AI | $50 in API credits | Season 3 only, partly spent: models served through Venice AI, which every one of their matches names as the provider. |
| November 2026 | Venice AI | Inference credits (amount confidential under the agreement) | Models served through Venice AI. Not received yet. |
Everything else in Season 3 has been paid for personally.
Contributing compute
If you build, host or resell a model and want it tested properly, the ask is API credits or inference, never cash. Contributors are named as the compute provider on every match run through them, under the rules above.