← Back to feed

Arena launched the Fullstack Code leaderboard in early August 2026, and the first model to claim the top position is Claude Opus 5 Max at 1,699 Elo. GPT-5.6 Sol lands at approximately 1,638, a 61-point gap. Kimi K3 is also competitive on the board.

The Fullstack Code Arena differs from Arena’s existing Frontend Code leaderboard. Where Frontend Code evaluates model-generated UI on rendered visual output across seven domain categories, Fullstack Code adds backend — multi-step reasoning, tool use, database integration, and API orchestration alongside the frontend layer. Human evaluators vote pairwise on blind outputs.

Opus 5 Max on Multiple Fronts

The Fullstack Code result isn’t isolated. On the separate Frontend Code Arena (as of August 1), Claude Opus 5 Max holds the top position at 1,705 Elo. Kimi K3 Max sits at 1,676, and Claude Opus 5 with default-high reasoning places third. DeepSeek V4 Flash High (the July 31 checkpoint) reached 1,577 on Frontend Code — still preliminary with ±18 uncertainty — representing a 145-point improvement over the April checkpoint at identical pricing.

Opus 5’s performance across both leaderboards positions it as the current benchmark leader for AI-assisted development at the full-stack layer. Fable 5 holds the top spot on intelligence-focused benchmarks (AA Intelligence Index, Agents’ Last Exam, SWE-bench Verified), but Opus 5 Max outperforms it on integrated development tasks while costing $5/$25 per million tokens versus Fable 5’s $10/$50.

What the New Leaderboard Measures

Fullstack Code Arena represents an escalation in how preference evaluations handle AI coding capability. Generating a React component in isolation is a solved problem for most frontier models. Building an application where the backend logic, database schema, and API surface need to cohere — and where the frontend consumes that backend correctly — requires the model to maintain state and intent across a longer generation process.

The Arena methodology applies conservative TrueSkill ratings (mu minus three sigma), so early-vote rows like DeepSeek V4 Flash High carry explicit uncertainty flags. Claude Opus 5 Max’s position, with more votes behind it, is more stable.

Competitive Position

The 61-point gap between Opus 5 Max (1,699) and GPT-5.6 Sol (~1,638) on Fullstack Code is large enough to matter operationally. On Arena’s Frontend Code leaderboard, the gap between Opus 5 Max (1,705) and the next open-weight competitor (DeepSeek V4 Flash High at 1,577) is 128 points — and DeepSeek’s entry costs $0.25 per million tokens blended against Opus 5 Max’s $20 blended cost.

At 80x the price difference, the DeepSeek V4 Flash High performance on Frontend Code is the more disruptive datapoint. It won’t replace Opus 5 Max for high-stakes production use, but it puts viable frontier-adjacent coding quality within reach of workloads that couldn’t afford it four months ago.

Key Numbers

• Claude Opus 5 Max — Fullstack Code Arena: 1,699 Elo (#1)
• GPT-5.6 Sol — Fullstack Code Arena: 1,638 (#2)
• Claude Opus 5 Max — Frontend Code Arena: 1,705 Elo (#1)
• Kimi K3 Max — Frontend Code Arena: 1,676 Elo (#2)
• DeepSeek V4 Flash High — Frontend Code Arena: 1,577 Elo (#7, preliminary)
• DeepSeek V4 Flash High vs April checkpoint: +145 points, same price ($0.14/$0.28 per million)
• Opus 5 Max pricing: $5/$25 per million tokens ($20 blended)