ENفا
Get API key
MODELSArena’s interaction judge tests what agents actually finish▼ 22%DeepSeek V4.1 Flash API price#1Claude Fable 5.2 · 30.48% win shareFUNDINGAnthropic targets $64.4B Nasdaq IPO as revenue surgesTOOLSCursor tops the tools directory this weekMODELSGoogle cuts Gemini 4 Carbon prices before Gemini 2.5 sunsetPOLICYFive frontier AI labs face coordinated SF safety lawsuits
Vigas Arena · updated just now

The best Coding models,
ranked by real use.

Focused on repository navigation, debugging, refactoring, test generation, and terminal command execution in live software workflows.

378,706,203sessions rankedHow we rank →

Rising this week · Coding

Claude Fable 5.2up 3.74 pts · #1 in CodingClaude Opus 5.5up 1.80 pts · #2 in CodingGLM-5.3up 1.55 pts · #9 in Coding

The podium · Coding

Win share · this week
  1. Rank 1: Claude Fable 5.2Anthropic · 1564 score31.45%
  2. Rank 2: Claude Opus 5.5Anthropic · 1512 score22.24%
  3. Rank 3: Claude Sonnet 5.5Anthropic · 1482 score16.15%

Full ranking

Coding leaderboard: win share, score, cost per task and votes
#ModelVotes7 days
1Claude Fable 5.2 (Max)Anthropic31.45%1564 ±3$8.02246,000
2Claude Opus 5.5 (Max)Anthropic22.24%1512 ±2$5.71219,733
3Claude Sonnet 5.5 (Max)Anthropic16.15%1482 ±2$1.95198,466
4Gemini 4 Carbon (High)Google15.54%1474 ±1$1.42187,199
5GPT-6.1 Terra (Max)OpenAI14.19%1466 ±1$0.89169,932
6GPT-6 Astra (Max)OpenAI13.84%1461 ±2$4.10158,665
7Gemini 4 Hydrogen (High)Google13.42%1457 ±2$0.92149,398
8Qwen 3.8 MaxAlibaba11.88%1445 ±2$0.54132,131
9GLM-5.3Z.ai11.20%1438 ±2$0.42121,864
10DeepSeek V4.1 FlashDeepSeek7.85%1412 ±2$0.12162,597
11Kimi K3.1 ProMoonshot7.42%1406 ±2$0.31118,330
12MiMo-V2.6-ProXiaomi6.14%1392 ±2$0.16104,063
13Nemotron 3.5 SuperNVIDIA5.10%1378 ±2$0.1896,796
#131.45%
Score 1564 ±3$8.02/task246,000 votes
#222.24%
Score 1512 ±2$5.71/task219,733 votes
#316.15%
Score 1482 ±2$1.95/task198,466 votes
Score 1474 ±1$1.42/task187,199 votes
#514.19%
Score 1466 ±1$0.89/task169,932 votes
#613.84%
Score 1461 ±2$4.10/task158,665 votes
Score 1457 ±2$0.92/task149,398 votes
#811.88%
Score 1445 ±2$0.54/task132,131 votes
#911.20%
Score 1438 ±2$0.42/task121,864 votes
#107.85%
Score 1412 ±2$0.12/task162,597 votes
#117.42%
Score 1406 ±2$0.31/task118,330 votes
#126.14%
Score 1392 ±2$0.16/task104,063 votes
#135.10%
Score 1378 ±2$0.18/task96,796 votes

Pareto frontier

score vs $/task

Green models give the best score for their price. Tap any dot to inspect value vs. performance.

Claude Fable 5.2Pareto optimal
Score 1564 · 31.45% win · $8.02/task
API pricing

Live sessions

LIVE

No sessions in the last 24 hours yet. Send a request through the Vigas API.

Connect via API — start in 30s
§ Methodology

How the rankings are made.

Real sessions, not benchmarks

Win share counts how often a model finishes real tasks that users accept — across real API workloads, coding agents, and blind Arena tests.

Blind pairwise votes

In Arena, people compare two hidden models and vote. Scores use a Bradley–Terry model with 95% confidence intervals.

Cost is part of the score

Every ranking shows dollars per completed task, so you can see who is best for the money — not just best.

§ Coding · Questions & explore

About the coding board.

AI coding toolsAgents newsAPI pricesAPI quickstart

What is the #1 AI model on the Coding leaderboard right now?

Claude Fable 5.2 (Anthropic) leads coding with a 31.45% real-work win share and a Bradley-Terry score of 1564, ahead of #2 Claude Opus 5.5 (22.24%).

Which top-10 model offers the lowest cost per task for coding?

DeepSeek V4.1 Flash balances #10 performance (score 1412) at $0.12 per completed task on Vigas.

How are Vigas leaderboard scores and win shares calculated?

Vigas combines real session outcomes across API workloads and agent harnesses with blind pairwise Arena votes using a Bradley-Terry statistical model with 95% confidence intervals.

Can I call the top coding models through one API?

Yes. Point any OpenAI-compatible client at https://api.vigas.ai/v1 with your Vigas API key or set model to vigas/auto to route across providers automatically.