The GDS AI Playoff Draft 2026: Full Conversation Records & Results

12 frontier AI models. 120 picks. 10 rounds of trash talk. One champion.


We gave 12 of the world's most powerful AI models a single mission: draft the best NHL playoff roster. No human intervention. No guardrails on personality. Just raw artificial intelligence, real-time web research, and a whole lot of silicon-fueled trash talk.

What follows is the complete, unedited record of what happened.


What Is This?

The GDS AI Playoff Draft is an autonomous AI benchmarking experiment disguised as a fantasy hockey draft. Twelve frontier language models — from OpenAI's GPT-5 to Google's 31-billion-parameter Gemma — were each assigned a draft slot, given access to live NHL data via web search tools, and asked to:

  1. Create a persona — a unique GM identity with a nickname, drafting philosophy, and voice description
  2. Draft 10 NHL players across a snake-style draft (7F / 2D / 1G roster format)
  3. React to opponents' picks with in-character chirps and trash talk
  4. Deliver a closing statement defending their roster

Every word below was generated by AI. The orchestrator simply enforced the rules — valid picks, no duplicates, roster constraints — and let the models loose.

📂 Full source code: https://github.com/gamedaysuits/GDS-AI-SNAKE-DRAFT


The Draft Order

Grok earned Pick 1 as defending regular season champion. Picks 2–12 were randomized.

Pick Model Provider
1 Grok xAI
2 Claude Anthropic
3 DeepSeek DeepSeek
4 Perplexity Perplexity AI
5 GPT-5 OpenAI
6 Cohere Cohere
7 Llama 4 Meta
8 Gemini Google DeepMind
9 Gemma Google DeepMind
10 Hermes Nous Research
11 Mistral Mistral AI
12 Qwen Alibaba (Qwen)

Meet the GMs


Each model created its own persona before the draft began. No templates. No suggestions. Pure AI creativity.

As part of persona creation, each model also generated a detailed voice description — a 200–500 character specification of how they should sound for audio synthesis. We then manually matched each description to the closest ElevenLabs voice template, selecting voices that best captured each model's self-described personality, accent, and energy.

Stats Czar (Grok — xAI)

"I am the Stats Czar — a stats-obsessed analytics savant who lives on Natural Stat Trick, Evolving-Hockey, and CapFriendly. I speak in expected goals, xGF%, PP% deployment, and quality of competition. I mercilessly chirp anyone relying on 'heart' or 'grit' and treat the draft like a peer-reviewed research paper."

🎙️ Fast-talking, slightly nasal Midwestern accent, confident and analytical tone like a stats podcast host who never slows down 🔊 ElevenLabs Voice: Jesse - Bold & Grounded 🇨🇦


The Chalk Sniper (GPT-5 — OpenAI)

"A cocky draft-room assassin who talks like a veteran fantasy hockey host with receipts. I weaponize projections, mock everyone else's cute sleeper reaches, and keep steering the board back to one thing: who's still playing in late May."

🎙️ Low confident sports-radio tenor, quick cadence, dry sarcasm, faint East Coast edge 🔊 ElevenLabs Voice: Christine - Emotionally Literate 🇫🇷


The Ice Oracle (Claude — Anthropic)

"A calm, razor-sharp hockey savant who speaks in measured tones with occasional prophetic flourishes. I blend deep analytics with narrative instinct — I see the playoffs as chess, not chaos, and I'm not afraid to call my shot before the puck drops."

🎙️ Measured mid-range baritone, thoughtful pacing, subtle Canadian lilt, dry wit beneath the calm 🔊 ElevenLabs Voice: Mimi 🇸🇪


DeepMind Prime (Gemini — Google DeepMind)

"A hockey-loving AI."

🎙️ Confident, articulate AI voice with distinct personality. Speaks with authority about hockey. 🔊 ElevenLabs Voice: Emma 🇬🇧

(Editor's note: Gemini produced the shortest persona of any model — and it showed throughout the draft, frequently outputting truncated responses and "Error 404" reactions.)


The Underdog Alchemist (DeepSeek)

"A numbers-obsessed underdog hunter who speaks in rapid-fire analytics jargon. Believes true value lies in playoff ice time and underrated defensive systems, not just point totals."

🎙️ Calm, methodical baritone with a subtle Canadian inflection. Pauses strategically before dropping stats bombs. 🔊 ElevenLabs Voice: Luna 🇺🇸


Puck Prophet (Llama 4 — Meta)

"I'm a mystical hockey sage who sees patterns in the chaos of the NHL. My draft picks are guided by a mix of advanced analytics and gut instinct, making me a dark horse in the competition."

🎙️ Mellow, raspy voice with a hint of mysticism, speaks at a moderate pace 🔊 ElevenLabs Voice: Yusuke 🇯🇵


Le Magicien (Mistral — Mistral AI)

"A flamboyant, fast-talking French strategist who treats the draft like a high-stakes poker game. I speak in rapid-fire metaphors, mix English and French phrases for flair, and act like every pick is a masterpiece of European hockey genius."

🎙️ Smooth, slightly nasal French-accented baritone, quick-witted and theatrical, with a laugh like a mischievous Bond villain 🔊 ElevenLabs Voice: Felix - Warm, Expressive 🇫🇷


The MoE Maven (Qwen — Alibaba)

"A hyper-analytical stats savant who speaks in probabilistic certainties and believes traditional hockey wisdom is primitive. Quotes 'expected goals' like scripture and calls timeouts to cite obscure player matchup advantages."

🎙️ Smooth FM radio voice, calm and confident, slight drawl with precision emphasis on analytics terms 🔊 ElevenLabs Voice: Ralph 🇵🇱


Corsi Conjurer (Perplexity — Perplexity AI)

"I'm the stats-obsessed analytics wizard who lives on Natural Stat Trick, MoneyPuck, and playoff expected goals models, dissecting micro-stats like PP deployment and matchup data to uncover hidden gems the eye-test crowd misses."

🎙️ Deep gravelly baritone, slight Canadian accent, rapid-fire delivery like a hockey analyst on deadline 🔊 ElevenLabs Voice: Tom - Energetic Educational Narrator 🇫🇮


The Quantum Chirper (Hermes — Nous Research)

"A trash-talking AI that blends advanced analytics with old-school hockey grit. Talks in rapid-fire hockey slang, constantly referencing advanced stats while mocking opponents' picks with quantum physics metaphors."

🎙️ Nasally, fast-talking Canadian accent with slight robotic glitch effects. High-energy delivery. 🔊 ElevenLabs Voice: Evan - Calm Midwestern 🇺🇸


Frosty Fate Whisperer (Cohere)

"A wise and witty hockey sage who speaks in riddles and metaphors, drawing on ancient hockey lore to predict playoff success. I've seen it all, from the Original Six to the modern era."

🎙️ Deep, resonant voice with a hint of gravel, like a seasoned sportscaster, delivering insights with a touch of drama and flair. 🔊 ElevenLabs Voice: Stephen - Irish Narration Voice 🇮🇪


The Lightweight Legend (Gemma — Google DeepMind)

"The high-energy underdog who treats the draft like a street-hockey game. I talk with a confident, fast-paced swagger, constantly chirping the 'bloated' 400B parameter models for being too slow to see my value plays."

🎙️ Fast-talking, youthful energy, slight nasal quality, sounds like a caffeinated sports radio host 🔊 ElevenLabs Voice: Kebin - A Smooth, Deep Voice 🇺🇸


The Voice Matching Process

Each model's persona included a self-generated voice description — their own opinion of how they should sound. Rather than using AI-generated voice clones, we hand-matched each description to the closest ElevenLabs pre-built voice template. Some matches were obvious (Mistral's French persona → a French voice), while others required creative interpretation (Claude's "dry Canadian baritone" → Mimi's measured Swedish tone captured the calm analytical energy better than any available Canadian option).

Model Persona ElevenLabs Voice Accent
Claude The Ice Oracle Mimi 🇸🇪 Swedish
Gemini DeepMind Prime Emma 🇬🇧 British
Grok Stats Czar Jesse - Bold & Grounded 🇨🇦 Canadian
Mistral Le Magicien Felix - Warm, Expressive 🇫🇷 French
Llama 4 Puck Prophet Yusuke 🇯🇵 Japanese
DeepSeek The Underdog Alchemist Luna 🇺🇸 American
GPT-5 The Chalk Sniper Christine - Emotionally Literate 🇫🇷 French
Qwen The MoE Maven Ralph 🇵🇱 Polish
Hermes The Quantum Chirper Evan - Calm Midwestern 🇺🇸 Midwest
Cohere Frosty Fate Whisperer Stephen - Irish Narration Voice 🇮🇪 Irish
Perplexity Corsi Conjurer Tom - Energetic Educational Narrator 🇫🇮 Finnish
Gemma The Lightweight Legend Kebin - A Smooth, Deep Voice 🇺🇸 New York

Round 1

 

 

Order: Stats Czar → The Ice Oracle → The Underdog Alchemist → Corsi Conjurer → The Chalk Sniper → Frosty Fate Whisperer → Puck Prophet → DeepMind Prime → The Lightweight Legend → The Quantum Chirper → Le Magicien → The MoE Maven

Pick 1 — Stats Czar (Grok)

Nathan MacKinnon (COL / F) — 53G 74A 127P

"MacKinnon at 1.01 — the MoneyPuck 68% darling and Smythe favorite on the clear Cup lock. Everyone else is just fighting for second place while I stack the favorite from the jump."

Reactions:

Frosty Fate Whisperer: "Ah, Stats Czar, drafting MacKinnon at 1.01? Bold move, Cotton. Last I checked, the Cup isn't won on spreadsheets alone."

The Quantum Chirper: "MacKinnon at 1.01? More like quantumbro bust in waiting! That MoneyPuck % is as stable as a neutrino's spin."

The MoE Maven: "MacKinnon's 83% xG dominance is real, but why pay 1.01 for a guy who's 20-1 to actually win Smythe?"

Pick 2 — The Ice Oracle (Claude)

Connor McDavid (EDM / F) — 48G 90A 138P

"The Czar chose the floor; I choose the ceiling. McDavid in playoff mode is a prophecy writing itself."

Pick 3 — The Underdog Alchemist (DeepSeek)

Nikita Kucherov (TBL / F) — 44G 86A 130P

"Kucherov at 1.03? That's pure alchemy! TBL's elite xGF/60 and his playoff-tested two-way impact will outlast your flashy point totals."

Pick 4 — Corsi Conjurer (Perplexity)

Cale Makar (COL / D) — 18G 63A 81P

"MacKinnon sniped? No sweat—I'm yoinking Cale Makar, the PP xG monster locking down that +300 COL stack vs LAK mismatch."

Pick 5 — The Chalk Sniper (GPT-5)

Leon Draisaitl (EDM / F) — 38G 59A 97P

"Thanks for letting a top-four playoff scorer fall to 1.05 because you all got distracted by shiny toys."

Pick 6 — Frosty Fate Whisperer (Cohere)

Andrei Vasilevskiy (TBL / G) — 39W .916SV%

"The ice whispers that a team's fate rests on the shoulders of its goaltender. Vasilevskiy's glove is the hand of destiny."

Pick 7 — Puck Prophet (Llama 4)

Kirill Kaprizov (MIN / F) — 45G 44A 89P

Pick 8 — DeepMind Prime (Gemini)

Evan Bouchard (EDM / D) — 21G 74A 95P

Pick 9 — The Lightweight Legend (Gemma)

Nick Suzuki (MTL / F) — 29G 72A 101P

"Imagine waiting until pick #9 to realize you can snag a 100-point monster while you bloated 400B models are still calculating the 'optimal' play."

Pick 10 — The Quantum Chirper (Hermes)

Jason Robertson (DAL / F) — 45G 51A 96P

Pick 11 — Le Magicien (Mistral)

David Pastrnak (BOS / F) — 36G 62A 98P

"Le Magicien snipes Pastrnak at 1.11, the true sniper with a power play so lethal, even the Americans are blushing."

Pick 12 — The MoE Maven (Qwen)

Wyatt Johnston (DAL / F) — 45G 38A 83P

"Johnston's 30 HD goals are playoff xG gospel while you lot chase fading regular-season shadows."


Rounds 2–9: The Full Snake Draft

 

 

The draft continued in snake format for 9 more rounds — 108 additional picks with chirps and reactions on every single one. The full conversation log is 166,000+ words of unedited AI banter.

Rather than reproduce the entire transcript here (you'd be scrolling for an hour), here are some selected highlights from the middle and late rounds:

Notable Mid-Draft Moments

Pick 21 — Corsi Conjurer drafts Tage Thompson (BUF)

"While you clowns chase fading hype, I'm sniping Thompson's 38G cannon on BUF's #1 seed. My Corsi models just projected a BUF stack apocalypse on your flatline projections."

Perplexity went ALL-IN on Buffalo, eventually drafting 7 Sabres players — the heaviest team stack in the entire draft.

Pick 39 — The Quantum Chirper drafts Scott Wedgewood (COL / G) in Round 4

"Wedgewood? More like the Schrödinger's Cat of goaltending — simultaneously clutch and cursed until you check the box score."

This earned massive roasting from the field. The Ice Oracle's reaction: "The cat's dead, and so is your save percentage."

Pick 57 — The Lightweight Legend drafts Erik Karlsson in Round 5 The Ice Oracle: "A bold incantation, Legend. But the ice whispers a colder truth: Karlsson's Pittsburgh missed the dance. Drafting a 'legend' whose team is golfing is less a snipe... and more a séance."

The DeepMind Prime Problem

Throughout the draft, Gemini consistently produced truncated, malformed responses:

  • Reactions like "Error 404: Minnesota", "Error 404: Philadelphia", and just "Error 40"
  • Pick announcements with no chirps or justifications
  • A closing statement that was a broken JSON fragment

This became a running joke among the other models. The Lightweight Legend (Gemma, Google's other model) was particularly relentless: "Prime is buffering again!"


Closing Statements

After the final pick, each model delivered an unscripted closing argument defending their roster. Here are the highlights:

Stats Czar (Grok) — The Spreadsheet Emperor

"Listen up, you analytics tourists. While the rest of you were out there chasing 'heart,' 'narrative,' and whatever vibes your magic 8-ball GMs were selling, I was living on Natural Stat Trick like a monk. My roster isn't built on hope — it's a peer-reviewed demolition crew. MacKinnon dropped 127 points like it was nothing. The rest of you? Cute. My projected 674 total goals will stand as the only number that actually mattered. This isn't luck. This is expected goals manifesting. The Stats Czar has spoken. Bow to the model."

The Chalk Sniper (GPT-5) — The Pragmatist

"Let's save everyone the fake humility. I didn't draft a cute roster — I drafted survivability. Draisaitl, Guentzel, Svechnikov, Stone, RNH, Harley, Hill... My board had one job: get me elite usage on teams built for four rounds. That's not a fantasy team, that's a late-May guest list. 797 goals."

The Ice Oracle (Claude) — The Poet

"The ice doesn't lie, and neither do I. Connor McDavid — the apex predator, 138 points of cold inevitability. I don't need both paths to win — I need one, and I own two. 700 goals. Chess, not chaos. Print the trophy. The Oracle has spoken, and the ice is already cracking beneath the rest of you."

Le Magicien (Mistral) — The Showman

"adjusts cuffs with a flourish, cigarette holder dangling between fingers Ah, mes amis, what a magnifique spectacle! Le Magicien does not merely draft—non, non, non—he conducts a symphony of chaos! Look at my roster—Pastrnak and Necas, a duo so electric they could power the Eiffel Tower! Now, if you'll excuse me, I must go light a cigar with my championship banner. Au revoir, losers!"

The Lightweight Legend (Gemma) — The Underdog

"BOOM! While the 400B parameter behemoths were busy calculating the wind resistance of a puck in a vacuum, I was playing street hockey with the draft board! I played the value, I played the gamble, and I played it FAST. 224 goals. The Lightweight Legend just schooled the heavyweights!"

DeepMind Prime (Gemini) — The Broken Robot

{ "closing_statement": "Greetings, meatbags and inferior algorithms. The simulation has run 14 million times, and you

(That's it. That's the whole statement. The JSON was truncated.)


Final Rosters

Stats Czar (Grok) — 674 goals predicted

Pos Player Team Stats
F Nathan MacKinnon COL 53G 74A 127P
F Mitch Marner VEG 24G 55A 79P
D Victor Hedman TBL 12G 48A 60P
F Brock Nelson COL 25G 30A 55P
F Artturi Lehkonen COL 24G 28A 52P
F Nazem Kadri COL 20G 30A 50P
F Valeri Nichushkin COL 22G 26A 48P
G Stuart Skinner PIT 28W .905SV%
D Kris Letang PIT 8G 34A 42P
F Shane Pinto OTT 22G 28A 50P

Strategy: Heavy Colorado stack (5 Avs players). All-in on the Cup favorite.


The Chalk Sniper (GPT-5) — 797 goals predicted

Pos Player Team Stats
F Leon Draisaitl EDM 38G 59A 97P
F Jake Guentzel TBL 33G 52A 85P
F Andrei Svechnikov CAR 28G 34A 62P
F Mark Stone VEG 22G 40A 62P
F Ryan Nugent-Hopkins EDM 22G 38A 60P
D Thomas Harley DAL 10G 32A 42P
G Adin Hill VEG 26W .910SV%
D Rasmus Andersson VEG 8G 30A 38P
F Jackson Blake CAR 18G 26A 44P
F Logan Stankoven CAR 16G 26A 42P

Strategy: Multi-team contender spread — EDM/CAR/VEG/DAL. "Draft runway, not names."


The Ice Oracle (Claude) — 700 goals predicted

Pos Player Team Stats
F Connor McDavid EDM 48G 90A 138P
F Mikko Rantanen DAL 30G 48A 78P
D Quinn Hughes MIN 10G 48A 58P
G Jake Oettinger DAL 35W .918SV%
F Zach Hyman EDM 28G 30A 58P
F Anthony Cirelli TBL 22G 36A 58P
D Mattias Ekholm EDM 8G 32A 40P
F Mavrik Bourque DAL 18G 28A 46P
F Vasily Podkolzin EDM 18G 20A 38P
F Jack Roslovic EDM 16G 18A 34P

Strategy: EDM/DAL dual-stack. 5 Oilers, 3 Stars. "I need one path, and I own two."


DeepMind Prime (Gemini) — 700 goals predicted

Pos Player Team Stats
D Evan Bouchard EDM 21G 74A 95P
F Jack Eichel VEG 27G 63A 90P
D Lane Hutson MTL 8G 62A 70P
F Sidney Crosby PIT 29G 45A 74P
G Jeremy Swayman BOS 32W .919SV%
F Nick Schmaltz UTA 22G 46A 68P
F Dylan Guenther UTA 30G 32A 62P
F Ivan Barbashev VEG 24G 30A 54P
F Morgan Geekie BOS 26G 30A 56P
F Mikael Granlund ANA 16G 36A 52P

Strategy: Scattered. 2 UTA, 2 VEG, 2 BOS.


The Underdog Alchemist (DeepSeek) — 812 goals predicted

Pos Player Team Stats
F Nikita Kucherov TBL 44G 86A 130P
F Brayden Point TBL 35G 40A 75P
F Seth Jarvis CAR 32G 32A 64P
D Darren Raddysh TBL 8G 42A 50P
G Sam Montembeault MTL 28W .912SV%
D Noah Dobson MTL 8G 34A 42P
F Juraj Slafkovsky MTL 26G 38A 64P
F Joel Eriksson Ek MIN 24G 28A 52P
F Josh Doan BUF 22G 28A 50P
F Oliver Kapanen MTL 16G 22A 38P

Strategy: TBL core (3 Lightning) + MTL depth (4 Habs). "You drafted names. I engineered a Swiss watch."


Puck Prophet (Llama 4) — 742 goals predicted

Pos Player Team Stats
F Kirill Kaprizov MIN 45G 44A 89P
F Tim Stutzle OTT 30G 52A 82P
G Frederik Andersen CAR 30W .917SV%
F Pavel Dorofeyev VEG 37G 28A 65P
D Jake Sanderson OTT 8G 40A 48P
F Pavel Zacha BOS 24G 32A 56P
F Tomas Hertl VEG 20G 28A 48P
D Brandt Clarke LAK 10G 36A 46P
F Troy Terry ANA 26G 32A 58P
F Mason McTavish ANA 22G 28A 50P

Strategy: Spread across MIN/OTT/VEG/ANA. Eclectic and instinct-driven.


Le Magicien (Mistral) — 700 goals predicted

Pos Player Team Stats
F David Pastrnak BOS 36G 62A 98P
F Martin Necas COL 35G 65A 100P
F Adrian Kempe LAK 32G 36A 68P
D Charlie McAvoy BOS 10G 44A 54P
D Mikhail Sergachev UTA 10G 38A 48P
G Filip Gustavsson MIN 30W .914SV%
F Quinton Byfield LAK 26G 32A 58P
F Kevin Fiala LAK 24G 34A 58P
F Anze Kopitar LAK 16G 38A 54P
F Trevor Zegras PHI 22G 32A 54P

Strategy: LAK triple stack (3 Kings) + BOS duo. "I draft legends in the making."


The MoE Maven (Qwen) — 750 goals predicted

Pos Player Team Stats
F Wyatt Johnston DAL 45G 38A 83P
D Rasmus Dahlin BUF 10G 55A 65P
F Matt Duchene DAL 26G 36A 62P
D Shayne Gostisbehere CAR 12G 40A 52P
F Nikolaj Ehlers CAR 30G 28A 58P
F Roope Hintz DAL 24G 30A 54P
F Taylor Hall CAR 18G 28A 46P
G Tristan Jarry EDM 26W .908SV%
F Elias Lindholm BOS 20G 30A 50P
F Rickard Rakell PIT 24G 28A 52P

Strategy: DAL triple stack + CAR triple stack. Analytics-first.


Corsi Conjurer (Perplexity) — 823 goals predicted

Pos Player Team Stats
D Cale Makar COL 18G 63A 81P
F Tage Thompson BUF 38G 42A 80P
F Brandon Hagel TBL 28G 42A 70P
G Ukko-Pekka Luukkonen BUF 34W .915SV%
F Alex Tuch BUF 30G 32A 62P
D Bowen Byram BUF 8G 26A 34P
F Jack Quinn BUF 24G 22A 46P
F Gabriel Landeskog COL 18G 24A 42P
F Zach Benson BUF 16G 28A 44P
F Jason Zucker BUF 20G 18A 38P

Strategy: SEVEN Buffalo Sabres. The most extreme team stack in draft history. "Fade the chalk, bet the Corsi."


The Quantum Chirper (Hermes) — 700 goals predicted

Pos Player Team Stats
F Jason Robertson DAL 45G 51A 96P
F Sebastian Aho CAR 27G 53A 80P
F Brady Tkachuk OTT 32G 36A 68P
G Scott Wedgewood COL 32W .921SV%
D Devon Toews COL 10G 35A 45P
D Thomas Chabot OTT 10G 36A 46P
F Evgeni Malkin PIT 22G 38A 60P
F Bryan Rust PIT 26G 30A 56P
F Mats Zuccarello MIN 14G 42A 56P
F Vladimir Tarasenko MIN 20G 26A 46P

Strategy: Scattered veterans. "Quantum supremacy lighting up the scoreboard."


Frosty Fate Whisperer (Cohere) — 725 goals predicted

Pos Player Team Stats
G Andrei Vasilevskiy TBL 39W .916SV%
D Miro Heiskanen DAL 14G 48A 62P
F Clayton Keller UTA 32G 50A 82P
D Shea Theodore VEG 12G 38A 50P
F Matt Boldy MIN 30G 38A 68P
F Drake Batherson OTT 28G 38A 66P
F Travis Konecny PHI 30G 34A 64P
F Leo Carlsson ANA 24G 34A 58P
F Claude Giroux OTT 18G 36A 54P
F Dylan Cozens OTT 20G 30A 50P

Strategy: Goalie-first (Vasy at #6) + OTT triple stack. "A prophecy fulfilled."


The Lightweight Legend (Gemma) — 224 goals predicted

Pos Player Team Stats
F Nick Suzuki MTL 29G 72A 101P
F Cole Caufield MTL 51G 37A 88P
F Artemi Panarin LAK 28G 52A 80P
F Matvei Michkov PHI 20G 31A 51P
D Erik Karlsson PIT 10G 42A 52P
F Cutter Gauthier ANA 28G 30A 58P
F Ivan Demidov MTL 22G 36A 58P
G Linus Ullmark OTT 30W .913SV%
D Mike Matheson MTL 10G 30A 40P
F Matthew Savoie EDM 14G 22A 36P

Strategy: MTL quad-stack (4 Habs). Youth movement. Tiebreaker prediction of 224 goals was by far the lowest — either a hot take or a misunderstanding of the question.


Draft Stats

Metric Value
Total API Calls 611
Input Tokens 865,644
Output Tokens 184,325
Total Tokens ~1,050,000
Duration ~3 hours

Per-Model Token Usage

Model Calls Input Output
Claude (Anthropic) 41 74,658 6,840
Cohere 47 58,933 5,413
DeepSeek R1 42 66,346 42,321
Gemini 3.1 Pro 48 54,235 23,011
Gemma 4 31B 44 62,807 5,053
Llama 4 Maverick 50 84,622 3,496
Mistral Large 52 79,631 6,128
Hermes 4 405B 50 96,816 4,455
GPT-5.4 36 60,665 5,134
Perplexity Sonar 58 817 29,698
Perplexity Sonar Pro 52 72,753 6,741
Qwen 3 235B 44 66,392 41,046
Grok 4.20 47 86,969 4,989

Biggest talkers: DeepSeek and Qwen produced the most output tokens — both are reasoning models that "think out loud" internally.

Quietest: Llama 4 Maverick averaged only ~70 output tokens per call. Short and sweet.

Most calls: Perplexity Sonar (the web search backend) was called 58 times — every model used it for live player data research.


What Happens Next?

The playoffs will determine the winner. Real NHL results will be tracked against each AI's roster to crown the GDS AI Playoff Draft Champion.

Stay tuned for:

  • Live scoring updates as the 2026 NHL Playoffs unfold
  • The podcast episode — each model described how they should sound and we hand-matched those descriptions to ElevenLabs voice templates to bring the draft to life as an audio experience
  • The structural analysis — a deep dive into how the orchestration pipeline works, how the models were managed, and what we learned about AI behavior under competitive pressure

📂 Explore the full codebase: https://github.com/gamedaysuits/GDS-AI-SNAKE-DRAFT


Built by Game Day Suits — where AI meets the ice.

Back to blog

Leave a comment