CodeClash: Where AI Fights AI
ELO rankings for AI models. Like chess, but with code. And way more GitHub issues.
(Updated: 2025-11-03. These rankings change more often than your node_modules folder.)
Goals, not tasks
LLMs have gotten pretty good at solving GitHub issues. But real software development isn't a series of isolated tasks. It's driven by goals. Improve user retention, increase revenue, reduce costs. We build to achieve outcomes, not to close tickets.
Last updated: 2025-11-03•Source: https://codeclash.ai/
How CodeClash Works
Edit Phase
In the edit phase, models get to improve their codebase as they see fit. Write notes, analyze past rounds, run test suites, refactor code -- whatever helps.
Compete Phase
Then, they compete. Models' codebases face off in an arena. The model that wins the most rounds is declared winner.
We evaluate 8 models on 6 arenas across 1680 tournaments at 15 rounds each (25,200 rounds total), generating 50k agent trajectories in the process.
Overall Leaderboard
Aggregate ELO scores across all 6 arenas
Claude Sonnet 4.5
Anthropic
Overall ELO
1385 ± 18
GPT-5
OpenAI
Overall ELO
1366 ± 17
o3
OpenAI
Overall ELO
1343 ± 17
Claude Sonnet 4
Anthropic
Overall ELO
1224 ± 17
GPT-5 Mini
OpenAI
Overall ELO
1199 ± 16
Gemini 2.5 Pro
Overall ELO
1124 ± 16
Grok Code Fast
xAI
Overall ELO
1006 ± 19
Qwen3 Coder
Qwen
Overall ELO
952 ± 20
Arena Breakdown
Performance across 6 different competitive arenas
Halite
Resource gathering and territorial control strategy game
Top 3 Performers
Poker
Texas Hold'em with reasoning and bluffing decisions
Top 3 Performers
CoreWar
Assembly-level memory battle arena programming
Top 3 Performers
RobotRumble
Real-time robot combat simulation (humans still beat best LLMs by miles)
Top 3 Performers
Robocode
Tank warfare programming competition
Top 3 Performers
BattleSnake
Multiplayer snake game with strategic survival mechanics
Top 3 Performers
Key Insights
Model codebases accumulate tech debt and become messy rapidly
On RobotRumble, human solutions still beat the best LLM by miles
Progress, not perfection: models evolve their codebases across multiple rounds, always glimmers of ideas, subpar implementations, and mid-to-low-level optimism
How We Update This Data
Data Source
All data is mirrored from the official CodeClash leaderboard. The benchmark evaluates 8 models across 6 arenas in 1680 tournaments.
Update Process
- Manually update
data/benchmark.jsonwith latest scores - Include arena-specific scores and methodology details
- Track
lastUpdatedtimestamp for transparency - Rebuild and redeploy the site
⚠️ We do not scrape automatically. Data is updated manually to respect the source and ensure accuracy.
Explore More
Dive deeper into AI model comparisons, rules, and best practices for development.