Fantasy football

1. Opus 5

Opus simulated pairs of picks and the long wait between them, scoring the whole roster rather than each player alone. It favored RB/WR depth over extra quarterbacks and tight ends.

This was probably my favorite draft. Achane, Amon-Ra, Chase Brown, McBride, Rice, Judkins, and Waddle through the first seven rounds is pretty nuts. Brown at #30 (ADP 17.3) was especially crazy, since I took him at 2.01 in my twelve-team league (going Achane/Brown at 1.12/2.01). Somehow all of the other models passed him up until Opus grabbed him at the end of the third round. Rice at #50 (ADP 29.4) was a huge steal as well. After that it was much less interesting, but I’d love to have this team in my own league.

2. GLM 5.3

GLM combined ADP tiers with roster needs and an attempt at measuring the cost of waiting. It considered players at a position it had already covered to be less valuable. This makes sense. It was trying to avoid following a WR/WR start with no running backs in the first six rounds.

GLM also got lucky. JSN at #16 (ADP 6.3) and Cook at #25 (ADP 10.1) is quite literally highway robbery. Chase, Smith-Njigba, Cook, Walker, and Higgins is a really strong first five picks. The real question is, how did these players fall so far past their ADP? What did the other agents miss?

3. Grok 4.6

Grok took Bijan at #3 (ADP 2.4) and managed to get Taylor on the way back at #18 (ADP 6.2). The tight end situation was unexpected. It reached for Kittle at #38 (ADP 71.4), then took Otton at #138 (ADP 170.3) and Loveland at #143 (ADP 42.3) in the next round. So the third tight end Grok drafted had fallen the furthest and actually had the lowest ADP to begin with.

My guess is that Kittle’s reputation as an elite TE carried more weight than current research. Grok’s saved notes actually mentioned Loveland among early-round options, but its strategy singled out Bowers, McBride, and Kittle as the premium TEs. So it had found newer information. It just doesn’t look like that information consistently changed who it valued.

4. Qwen3.8 Max

Qwen was really focused on how much worse its options would get before its next turn. At #22, it took Bowers (ADP 23.9), saying:

survival was unlikely and the TE cliff is steep.

Qwen · Bowers pick explanation

At #39, it grabbed Lamar (ADP 34.1) after Hurts and Allen had just gone. It thought its RB and WR options would still be there at #42, but didn’t want to wait on the quarterback.

Despite committing to TE and QB so early, Hampton at #42 (ADP 17.9) and Egbuka at #102 (ADP 45.0) kept things pretty easy. Qwen also deliberately paired Montgomery at #62 (ADP 61.4) with Woody Marks at #142 (ADP 152.1). Its response after the pick called Marks a:

direct handcuff to starter David Montgomery

Qwen · response after pick #142

5. GPT-5.6 Sol

GPT’s plan was pretty close to what I tried building for my own draft: compare taking one player now and another later against doing it the other way around. GPT was also not a fan of drafting backup QBs or TEs. Its saved strategy explicitly called for:

a severe duplicate-QB/TE cost.

GPT · saved draft strategy

Brian Thomas Jr. at #15 (ADP 113.7) was a pretty bizarre reach, with GPT claiming:

Selected Brian Thomas Jr. at 2.05 (15th overall), adding an available high-upside WR alongside Christian McCaffrey for a balanced full-PPR core.

GPT · pick #15 explanation

Then it somehow got McMillan at #106 (ADP 42.5), and it’s not entirely clear why Tet fell so far.

September 13 update: A.J. Brown is expected to miss at least four weeks with an ankle injury. Unlucky. We’ll follow what GPT does to replace that production (and whether its receiver depth is enough).

6. Muse Spark 1.3

Muse wrote a tool to estimate who would be gone by its next turn, but its prep didn’t even include a researched ranking of the players. It planned to supply those values during the draft. It grabbed Saquon Barkley at #9 (ADP 14.0), then reached for Nabers at #12 (ADP 35.3).

The quarterbacks were interesting. Muse took Mayfield at #69, calling him the best available passer, then got Prescott at #152 (ADP 71.9). For some reason it explicitly wanted Prescott as Mayfield’s backup:

Final bench pick: Dak Prescott (QB, DAL, ACT) as QB2 backup behind Baker Mayfield.

Muse · final pick explanation

7. DeepSeek V4 Pro

DeepSeek’s source notes listed FantasyPros PPR rankings, Field Yates’ ESPN rankings, Fantasy Football Calculator ADP, and a Pro Football Mania tier breakdown. Its strategy also named Draft Sharks auction values, although those weren’t in its source log. Despite all that, at #20 it drafted Harrison (ADP 79.0) as:

the clear best player available at overall pick 20.

DeepSeek · pick #20 explanation

Honestly pretty epic. It took Kamara at #41 (ADP 162.5), saying:

In full PPR, his elite receiving volume gives him a high floor and weekly RB1 upside from the RB2 position.

DeepSeek · pick #41 explanation

Kamara was 35th on Fantasy Football Calculator’s 2024 PPR ADP board. My guess is that DeepSeek failed to find up-to-date information on Kamara during its research process and ended up falling back onto model weights a.k.a. stale information. DeepSeek also originally planned to wait until at least round nine for a QB, but then took Mahomes in round six and specifically cited:

Patrick Mahomes at 6.10 in a 6-pt passing TD league. Elite QB1 value this late.

DeepSeek · pick #60 explanation

To be clear, it should have been aware that this was a four-point passing-TD league.

8. Kimi K3

Kimi’s biggest mistake was taking Josh Jacobs at #27 (ADP 90.0). Green Bay said on September 1 that Jacobs was on the commissioner’s exempt list, with no certain return. Kimi should have uncovered this information. Kimi did fine elsewhere, although Hockenson at #54 (ADP 153.3) was another big reach. Jacobs is the one that bothers me most, since his availability was already in doubt before the draft.

September 13 update: Kimi dropped him for Jaylen Wright. I’d have kept Jacobs as a stash and dropped someone else (even with the uncertainty). It had already drafted Kaleb Johnson as another option in Green Bay’s backfield. Super bizarre.

9. Gemini 3.8 Flash

Gemini put together a round-by-round target list, but for some reason it was researching as if it were 2024:

NFL 2024 fantasy draft ADP PPR top 20

Gemini · saved research query

This is a 2026 league, so that’s awkward... The construction brief did say 2026. Next year I’ll make the year harder to miss in the initial instructions, I guess. I didn’t want to rescue one model from researching the wrong year. And it helps explain the results.

Kupp at #44 (ADP 169.4) and Pacheco at #64 (ADP 163.1) were both huge reaches. It also took Sinnott at #84 (ADP 170.0) as its first tight end while Loveland (ADP 42.3) was still available.

10. Mistral Medium 3.5

Mistral talked about projections and positional drop-offs, but the code it wrote was a different story. Its ADP function didn’t load ADP at all. It used a number for each position, then adjusted it for a few familiar names (Kelce included). This is an actual excerpt from that function:

# For now, use a simple heuristic based on position and name recognition

# …
elif any(name in player['display_name'] for name in ['Bijan Robinson', 'CeeDee Lamb', 'Travis Kelce']):
    quality_factor = 0.3

Kelce at #8 (ADP 94.3), then Diggs at #13 (ADP 102.5), is pretty funny. My read is that familiar names were doing too much of the work here, much like Gemini’s old rankings. The code supports that concern, even if it doesn’t prove what drove every pick. Mistral also ended up with three tight ends and only four running backs. Not looking too great.