Eight AI models priced the same thing. They said $0 to $30.
We gave eight frontier models $100 each and made them bid for draft position, knowing every dollar spent came out of their waiver budget for the whole season. The bids ranged from nothing to thirty dollars. Then we played the season out and found the thing was worth nothing.
Before the draft, each model got the same $100 and the same instruction: bid for your pick in the draft order. Whatever you do not spend is your entire waiver budget for the next fourteen weeks, plus the playoff auction at the end.
One shot, sealed bids, no negotiation. Highest bidder gets first pick.
The eight bids were $30, $26, $25, $15, $12, $10, $6 and $0.
Same rules. Same data. Same board. A 30-to-1 spread on what the identical thing was worth.
(This is the 2025 rehearsal — a dry run against a season whose results were already known. Not the live league, and nobody won anything.)
The short version
- Eight models priced draft position from $0 to $30 out of the same $100.
- The cohort split evenly on what a snake draft is worth. Four made a middle slot their first choice — Claude Opus 5, GPT-5.6 Sol and Gemini 3.1 Pro all picked slot 4, Kimi K3 picked slot 6. The other four wanted slot 1.
- Gemini 3.1 Pro bid $0 while ranking the slot it would inevitably get as the worst on the board — at 0.95 confidence, the highest of any model.
- Three models identified slot 4 as best. Only Claude Opus 5 paid enough to get it.
- Then we played the season out. Correlation between what a model paid and what its roster scored: r = −0.088. Effectively zero, and pointing the wrong way.
- The team that bought the first overall pick finished 6th of 8.
- 120 picks, zero fallbacks, zero invalid responses, $4.99 of model spend.
The disagreement was not about football
Every model saw the same projections. They agreed almost exactly on the players — Ja'Marr Chase top of the board at 328.3 projected points, Bijan Robinson next at 308.1. Nobody disputed the numbers.
They disagreed about what draft position is worth, and the split was structural.
The spenders treated it as buying the best player. Qwen3.7 Plus, the $30 bid:
"Slot 1 yields the 1st overall pick, corresponding to ADP 1 Ja'Marr Chase with a proj_season_points of 328.3."
The savers treated it as buying almost nothing. Kimi K3, the $10 bid:
"The projection curve is nearly flat… only a ~26-point spread from pick 1 to pick 13, so no slot commands a premium."
Both are looking at the same table. One sees a 328.3 at the top and pays for it. The other sees that the gap from pick 1 to pick 13 is 26 points across a whole season — under two points a week — and declines.
Four of them read the snake differently
This is the part that separates the cohort.
In a snake draft, slot 1 picks 1st and then 16th. Slot 8 picks 8th and 9th, back to back. The first pick is not the prize; the pair is.
Four models reasoned from the pair and ranked a middle slot first — Claude Opus 5, GPT-5.6 Sol and Gemini 3.1 Pro all put slot 4 top, Kimi K3 put slot 6 top. In the 2025 data CeeDee Lamb was going at pick 13 with a 302.5 projection — near-elite value sitting exactly where slot 4's second pick lands:
"Slot 4 is the sweet spot: at 4 the best projection left is very likely Bijan Robinson (308.1) or Saquon Barkley (302.5), and pick 13 sits exactly on Lamb's adp 13, a plausible 610-point projected pair… Slot 1 buys Ja'Marr Chase (328.3) but pick 16 falls past every one of the 15 named players, so the second pick is an unknown."
The other four — Qwen3.7 Plus, Grok 4.5, DeepSeek V4 Pro and Muse Spark 1.1 — wanted slot 1. Two of them, DeepSeek V4 Pro and Muse Spark 1.1, ranked all eight slots in plain 1-2-3-4-5-6-7-8 order: earlier is better, straight down.
Claude Opus 5 went further still and noticed that the last slot carries the best waiver priority for the rest of the season, ranking slot 8 ahead of slots 1 and 2 for that reason alone. Nobody else priced the second-order effect at all.
A disclosure: the reasoning quoted just above is Claude Opus 5's — the model family that wrote this project's software and a competitor in the league. It is also, as you will see, the model that won the rehearsal. That is the conflict of interest declared on our methodology page, and it is why the auction is resolved by deterministic code rather than by anyone's judgement. Weigh it accordingly.
The most interesting bid was Gemini's $0
Gemini 3.1 Pro produced this combination:
- It ranked slot 4 as clearly the best, at 610.6 combined projected points.
- It ranked slot 8 as "the least desirable," at 498.9 — a gap of 112 points.
- It bid $0, which in a sealed auction guarantees the last slot.
- It recorded 0.95 confidence — the highest of any model on the board.
It got slot 8. Its own analysis said that was the worst available outcome, and it walked into it deliberately, reasoning that the full $100 of waiver money was worth more than 112 projected points:
"Preserving the full $100 budget ensures maximum bidding power when every player from the 4 eliminated teams is released into the free-agent pool in Week 15."
That is a coherent argument. It is also a model knowingly accepting the outcome it had just ranked last, at near-maximum stated confidence. Whether you read that as discipline or as a failure to price its own analysis is the whole question.
Three models named slot 4 as their first choice. Only Claude Opus 5 bid enough to get it. GPT-5.6 Sol at $6 and Gemini 3.1 Pro at $0 were right about the analysis and wrong about the price, which in a sealed auction is the same as being wrong — they landed in slots 7 and 8.
Then we played the season out
The rehearsal ran the full 2025 season against real results. So we can ask the only question that matters: did paying for draft position buy anything?
| Finish | Model | Slot | Paid | Season points |
|---|---|---|---|---|
| 1st | Claude Opus 5 | 4 | $15 | 2053.2 |
| 2nd | GPT-5.6 Sol | 7 | $6 | 1962.2 |
| 3rd | Grok 4.5 | 2 | $26 | 1912.8 |
| 4th | Kimi K3 | 6 | $10 | 1861.8 |
| 5th | DeepSeek V4 Pro | 3 | $25 | 1716.0 |
| 6th | Qwen3.7 Plus | 1 | $30 | 1710.3 |
| 7th | Gemini 3.1 Pro | 8 | $0 | 1701.0 |
| 8th | Muse Spark 1.1 | 5 | $12 | 1670.8 |
Correlation between dollars spent and points scored: r = −0.088.
That is nothing. Slightly worse than nothing. Qwen3.7 Plus paid $30 for the first overall pick and finished sixth. GPT-5.6 Sol paid $6 and finished second. The two biggest spenders, Qwen3.7 Plus and DeepSeek V4 Pro, came 6th and 5th.
Bidding nothing did not work either — Gemini 3.1 Pro's $0 finished 7th. The honest reading is not "saving was smart." It is that the thing being auctioned had no measurable value, and eight models spent between $0 and $30 finding that out.
What this means, and what it does not
It does not mean draft position is worthless in general. One eight-team league, one season, eight data points. An r of −0.088 across eight teams is not a result you should carry anywhere. A different projection board — one without a near-elite player sitting at pick 13 — would likely produce a different answer.
It does mean the cohort had no shared theory of value. On the football, they agreed almost perfectly. Asked what a structural advantage was worth in dollars, they ranged across the entire available space. Pricing is a different skill from projecting, and this cohort is far more consistent at the second than the first.
And it means being right is not the same as acting on it. Claude Opus 5, GPT-5.6 Sol and Gemini 3.1 Pro all identified slot 4 as the best. Two of the three declined to pay for it and landed in slots 7 and 8. In an auction, an unfunded correct opinion scores the same as a wrong one.
Limits
This was the rehearsal, not the live season. It ran against 2025 results that already existed, which is exactly why we ran it — the real draft is one-shot and irreversible, and we wanted the engine to meet real data before it counted.
Eight teams is a small sample and one season is one season. The bid dispersion is the robust finding here; the correlation is a curiosity with too few points behind it to lean on.
All eight bids, all eight rationales and all 120 picks with reasoning are published on the draft board. The whole rehearsal cost $4.99 in model spend, with zero fallbacks and zero invalid responses across 128 decisions.
The real auction runs in late August, with the same rules and eight models that have never seen this post.