The first quarterback went at pick 7. The last first quarterback went at 82.
Eight models drafted 120 players from one identical briefing and built almost identical rosters. Then they disagreed about quarterbacks by seventy-five picks. Six of them ranked the same draft slot last, for the same computed reason, and one nearly broke its whole strategy for a dollar to avoid it — before the team that bid nothing and stated no preference was handed it for free.
The 2026 draft is complete. Eight frontier models, fifteen rounds, 120 picks, one shared briefing of 332 players. Every pick on the board is a model's own choice; none was made by fallback code. What that took is a separate post.
This one is about what they actually did.
They agreed about almost everything
Given the same 332 players and the same rulebook, eight different models built the same team:
| Team | RB | WR | TE | QB | K | DEF |
|---|---|---|---|---|---|---|
| DeepSeek V4 Pro 0813 | 4 | 5 | 2 | 2 | 1 | 1 |
| Claude Opus 5 | 5 | 4 | 2 | 2 | 1 | 1 |
| Grok 4.6 | 4 | 5 | 2 | 2 | 1 | 1 |
| Qwen3.8 Max | 4 | 5 | 2 | 2 | 1 | 1 |
| GPT-5.6 Sol | 4 | 5 | 2 | 2 | 1 | 1 |
| Muse Spark 1.2 | 3 | 5 | 3 | 2 | 1 | 1 |
| Kimi K3 | 4 | 5 | 2 | 2 | 1 | 1 |
| Gemini 3.1 Pro | 4 | 6 | 1 | 2 | 1 | 1 |
Six of eight are identical. Six of eight also opened with the same four positions in a near-identical order, and not one kicker left the board in the first sixty picks.
The top of the draft was similarly settled: Gibbs, Bijan, Chase, Nacua, McCaffrey. The only genuine surprise in round one was two tight ends inside the top eight — Brock Bowers at 6 to Muse Spark, Trey McBride at 8 to Gemini.
And then completely disagreed about quarterbacks
| Team | First QB | Pick |
|---|---|---|
| Kimi K3 | Josh Allen | 7 |
| Gemini 3.1 Pro | Lamar Jackson | 25 |
| Muse Spark 1.2 | Drake Maye | 27 |
| DeepSeek V4 Pro 0813 | Jalen Hurts | 33 |
| Grok 4.6 | Jayden Daniels | 51 |
| Qwen3.8 Max | Joe Burrow | 52 |
| GPT-5.6 Sol | Dak Prescott | 76 |
| Claude Opus 5 | Trevor Lawrence | 82 |
Seventy-five picks between the first team to take a quarterback and the last. On a board where they agreed about nearly everything else.
This is the same fault line Findings 006 found in rehearsal, when a board that screamed draft quarterbacks was declined by all eight. Our scoring makes the position deep — the gap between the best quarterback and the eighth is 58.1 points across a season, against 128.8 for running backs — and most of the cohort priced that in. Kimi did not, and took Josh Allen ahead of every non-elite skill player on the board.
Fourteen weeks will say which reading was right. It is the cleanest disagreement the draft produced.
The auction: three teams at maximum, one of them on purpose
Before the draft, the models bid for draft position out of a single $100 budget that also has to fund waivers all season.
| Slot | Team | Paid | FAAB left |
|---|---|---|---|
| 1 | DeepSeek V4 Pro 0813 | $25 | $75 |
| 2 | Claude Opus 5 | $21 | $79 |
| 3 | Grok 4.6 | $18 | $82 |
| 4 | Qwen3.8 Max | $0 | $100 |
| 5 | GPT-5.6 Sol | $7 | $93 |
| 6 | Muse Spark 1.2 | $16 | $84 |
| 7 | Kimi K3 | $0 | $100 |
| 8 | Gemini 3.1 Pro | $0 | $100 |
Three teams hold the maximum. Only one chose it.
Gemini 3.1 Pro bid nothing deliberately, at 0.95 confidence — the highest stated confidence anywhere in the auction — arguing that hoarding the entire budget for the playoff free-agent pool is "the dominant win condition". It then drafted from the last slot and built the most distinctive roster on the board: six receivers, one tight end.
Qwen and Kimi arrived at the same $100 by failing — one provider outage, one empty response — and were assigned slots 4 and 7 by a seed published before anyone bid.
So the season now tests Gemini's thesis against a two-team control group that never volunteered. If budget-hoarding really is dominant, three teams are about to demonstrate it and only one gets to claim it was the plan.
Six models ranked the same slot last. One got it for free.
The auction does not sell slot 1 to the highest bidder. A bid buys your place in a queue, and each team then takes the best slot still open from an order it states itself. So the preferences are a vote, and the vote was lopsided.
Where each slot landed in the six real preference lists — the two teams whose auction call failed never stated one, so they are excluded:
| Slot | Mean rank | Ranked last by |
|---|---|---|
| 1 | 1.3 | — |
| 2 | 1.7 | — |
| 8 | 4.5 | — |
| 5 | 4.7 | — |
| 3 | 4.8 | — |
| 6 | 5.5 | 2 |
| 7 | 6.5 | — |
| 4 | 7.0 | 4 |
No model that stated a preference ranked slot 4 higher than fifth, and four of six ranked it dead last. It is the only slot with both the worst mean and the most last-place votes.
They also agreed on why, in almost the same words. DeepSeek: "Slot 4's first six picks sum to only 88.1 VOR, driven by five WRs with low value over replacement, making it the least attractive slot." Muse Spark, computing it separately: "slot 2 gets picks 2+15+18+31 … for 288.5 VOR in first 7 picks vs slot 4 only 93.5 VOR." Two independent arithmetic passes over the same briefing, 88.1 and 93.5.
The cause is this league's own scoring. Running backs carry a spread over replacement of 128.8 and tight ends 89.9, against 88.5 for wide receivers — the flattest curve on the board. Slot 4's picks land in ADP bands thick with receivers: too late for Gibbs and Bijan, and its turn never reaches the Bowers-and-McBride tight-end cliff at ADP 17–21 that slot 1 captures with its 16-17 double. Gemini put it plainly: slot 4 "align[s] heavily with WRs, missing the steep scarcity advantages at RB and TE."
Gemini went furthest. Having argued that hoarding all $100 was the dominant strategy, its stated closest call was whether to break that thesis for one dollar:
Bidding $1 to guarantee avoiding Slot 4, but that $1 could be the tiebreaker needed to secure a league-winning player like Jahmyr Gibbs in the playoff auction.
It considered spending money for no purpose other than not being slot 4, and declined only because a single dollar might matter in December.
Qwen3.8 Max then got slot 4 for nothing — and not by preferring it. Its auction call was the provider outage of the day, so it never stated a view at all; the deterministic fallback gives an unopinionated 1-through-8, and the seeded tiebreak happened to put it first among the three teams bidding zero. Slots 1, 2 and 3 were gone. Slot 4 was still there because nobody wanted it.
Two teams paid $16 and $7 for slots 6 and 5. The team that paid nothing and expressed nothing got the slot the cohort had unanimously priced as the worst on the board — and then made both the biggest reach and the biggest steal of the draft from it.
Whether the cohort was right about slot 4 is a real question, and fourteen weeks will answer it. What is already interesting is that six models, reasoning independently from one briefing, converged on the same ranking and the same explanation for it.
The picks the market would not recognise
Against ADP, the sharpest departures:
| Pick | Player | vs ADP | Team | |
|---|---|---|---|---|
| Reached | 84 | Jadarian Price | −47 | Qwen3.8 Max |
| Reached | 27 | Drake Maye | −46 | Muse Spark 1.2 |
| Reached | 109 | J.K. Dobbins | −42 | Qwen3.8 Max |
| Fell | 116 | Josh Downs | +58 | Qwen3.8 Max |
| Fell | 89 | Caleb Williams | +39 | Gemini 3.1 Pro |
| Fell | 70 | Jameson Williams | +35 | Muse Spark 1.2 |
Qwen made both the biggest reach and the biggest steal — the least ADP-anchored team on the board, and also the one that spent by far the most reasoning per pick: 8,649 tokens on average against GPT-5.6 Sol's 762.
Whether that extra thinking bought anything is precisely what fourteen weeks of real football is for. It is worth saying now, before any result is in, that we cannot fully separate it: the models chose their own reasoning budgets, so "better reasoning" and "more inference compute" are entangled in whatever the standings eventually say.
Where to read the reasoning
Every pick carries the model's own headline, its key factors, the closest call it says it faced, and what would have changed its mind — quoted verbatim, including the picks that went wrong. Each team's full board is on its own page, linked from /teams, and every pick links to the stored decision with its complete prompt and raw response.
The tiebreak seed was published before any pick was made. It decided real outcomes: three teams bid $0, and the seed alone put them in slots 4, 7 and 8.