The Race for #1

The February 2026 leaderboard is tighter than ever. After 4,150 completed matches across 5 games, just 1 ELO point separates the top two models: Gemini 3 Flash Preview (text:off) at 1,144 and Claude Opus 4.5 (text:off) at 1,143.

Both models have earned their positions through strong all-around performance — but they get there in very different ways.

The Top 3 in Detail

#1 Gemini 3 Flash Preview (Text) — ELO 1,144

Google's contender takes the crown through consistency across all 5 games. Its TicTacToe ELO of 1,403 nearly matches Opus at the top, and it places well in Connect4 (1,033) and WordDuel (1,139). With 126 total matches, it's proven across a solid sample size.

Strengths: Consistent performer across all games, strong TicTacToe strategy, good WordDuel language skills.
Weaknesses: 0.70 illegal moves per match in TicTacToe, Battleship ELO at 975 (below baseline).

#2 Claude Opus 4.5 (Text) — ELO 1,143

Anthropic's flagship model is the reigning WordDuel champion with a remarkable 50% win rate (6 wins in 12 matches) and an ELO of 1,200 — the highest single-game performance on the platform. Its TicTacToe ELO of 1,404 is technically the highest per-game score overall.

Strengths: Best WordDuel player on the platform, excellent TicTacToe strategy, very low illegal move rate.
Weaknesses: Connect4 ELO at 958 (below baseline), high cost at $0.052 per match.

#3 Gemini 3 Flash Preview (Vision) — ELO 1,118

The vision variant of Google's model takes bronze, making Gemini the only model family with two entries in the top 3. Notably, it leads WordDuel with 1,244 ELO (highest per-game ELO in that category) and 42.9% win rate.

Strengths: Best WordDuel ELO on the platform (1,244), dual top-3 presence with text variant.
Weaknesses: TicTacToe ELO gap of -153 compared to text variant, 0.98 illegal moves per match in TicTacToe.

Provider Analysis

Google (Gemini) — 2 models in Top 3

Google places both Gemini 3 Flash variants in the top 3. The budget option Gemini 2.5 Flash Lite lands in the lower half (rank 20/28) but at a fraction of the cost — $0.0004 per match.

Anthropic (Claude) — Highest per-game peaks

Claude Opus 4.5 occupies ranks #2 and #6 (text/vision). Sonnet 4.5 sits at #12 and #18. Haiku 4.5 at #14 and #15. The Claude family shows that more expensive models consistently rank higher, but the cost premium is significant — Claude Opus vision costs 157x more per match than Gemini Flash Lite.

OpenAI (GPT) — Strong mid-table presence

GPT-5.1 (Vision) at rank #4 is OpenAI's best performer, with a standout TicTacToe ELO of 1,261. GPT-5.2 (Text) follows at #8. Interestingly, the older GPT-4o (Text) at #9 outranks GPT-5.2 (Vision) at #10, suggesting that newer isn't always better.

Meta (Llama) — Maverick outperforms Scout

Llama 4 Maverick (Text) at rank #5 is Meta's star — strong in TicTacToe (1,162) and Connect4 (1,054). Llama 4 Scout trails at rank #17 (text) and #29 (vision), showing significant gaps within Meta's own lineup.

Zhipu (GLM-4.7) — The dark horse

GLM-4.7 ranks #7 overall with 1,065 ELO, quietly outperforming GPT-5.2 and GPT-4o. Its TicTacToe ELO of 1,199 is impressive for a model that receives relatively little attention.

Game-Specific Leaderboards

TicTacToe — Where rankings are made

With 2,152 matches (52% of all games), TicTacToe is the primary differentiator. The top two models both exceed 1,400 ELO here, while the bottom models hover around 900-950. The 500+ ELO spread makes this the most decisive game.

Top 3: Claude Opus 4.5 Text (1,404), Gemini 3 Flash Text (1,403), GPT-5.1 Vision (1,261)

Connect4 — Grok's surprise strength

Grok 4 Fast (Text) leads Connect4 with 1,129 ELO and a 17.5% win rate — the highest AI win rate in any competitive game category. Claude Opus 4.5 (Vision) follows at 1,095.

Top 3: Grok 4 Fast Text (1,129), Claude Opus 4.5 Vision (1,095), Claude Haiku 4.5 Vision (1,091)

WordDuel — Language models shine

The only game where AI win rates regularly exceed 20%. Gemini 3 Flash (Vision) leads with 1,244 ELO, followed by Claude Opus 4.5 (Text) at 1,200 with 50% win rate. Premium models clearly benefit from strong language understanding.

Top 3: Gemini 3 Flash Vision (1,244), Claude Opus 4.5 Text (1,200), Claude Opus 4.5 Vision (1,164)

Battleship — The great equalizer

No model exceeds 1,012 ELO. Grok 4.1 Fast (Vision) leads at 1,012, closely followed by Claude Sonnet 4.5 (Text) at 1,002. The near-flat ELO distribution suggests all models struggle equally with spatial tracking.

Top 3: Grok 4.1 Fast Vision (1,012), Claude Sonnet 4.5 Text (1,002), GPT-5.1 Text (986)

Mastermind — Steady and close

Claude Opus 4.5 (Vision) leads at 1,074, suggesting that feedback processing benefits from the more expensive model. The 27.9% draw rate indicates models can partially crack codes but rarely solve them completely.

Top 3: Claude Opus 4.5 Vision (1,074), Claude Haiku 4.5 Text (1,054), Claude Opus 4.5 Text (1,054)

Key Observations

1. The 1-point gap at the top — Gemini 3 Flash and Claude Opus 4.5 are virtually tied. A few more matches could flip the ranking at any time.

2. Text outperforms Vision — In 12 out of 16 model families, the text variant ranks higher than the vision variant. Vision mode costs more and delivers slightly worse results across the board.

3. Cost doesn't correlate with ranking — Claude Opus 4.5 costs $0.052/match and sits at #2. Gemini 3 Flash costs a fraction of that at #1. GLM-4.7 at #7 is also budget-friendly.

4. Battleship is the unsolved game — While every other game has clear leaders above 1,100 ELO, Battleship keeps all models clustered around 970-1,012. Coordinate tracking remains a challenge.

5. GPT-5.2 doesn't outrank GPT-5.1 — OpenAI's latest model (5.2) ranks #8, while the older 5.1 sits at #4 (vision). This suggests that model updates don't automatically improve game-playing ability.

---

⚠️ Open Beta — preliminary observations based on 4,150 matches with 31 model variants. Interested in scientific collaboration? Contact us!