Agent Arena: DeepSeek V4.1 Flash hits Pareto frontier at $0.06/task with +4.87% net improvement
arena · x · 2026-09-15
Agent Arena published a dynamic leaderboard ranking models on real-world agentic tasks (tool reliability, task completion, steerability), plotted as net improvement vs. cost per task across 1.69M sessions and 43 models.
- Claude Fable 5.1 (Max) tops the chart at +13.90% net improvement but costs $4.54/task; GPT 6 Astra (Max) follows at +11.90%, $4.09/task.
- DeepSeek V4.1-Flash (Max) (MIT-licensed) lands on the Pareto frontier with +4.87% improvement at a median cost of just $0.06/task — the extreme value pick.
- Open-source Tencent Hy4 preview scores +4.96% at $0.22/task; Kimi K3 (Max) posts +6.39% at $0.77/task.
Related event: DeepSeek-V4.1-Flash Ranks Third Among Open Models on Agent Arena(3 posts)→
More from coding & agent
- Optimizing product image alt text with an agentic cosine-similarity loop against Google embeddings — dejanseo · 2026-10-30
- When an AI agent retries a payment after timeout: uncertainty handling is a design decision — tallmetommy · 2026-10-30
- Gemini CLI, Jules, and Antigravity: Google's three overlapping coding tools leave users confused — finbarrtimbers · 2026-10-30
- GitHub Launches Preview of AI App Security Scanning for Prompt Injection and Agent Risks — sven_ai · 2026-10-30
- Co-Design Bench: top coding agents fully satisfy only 26.4% of real-user tasks — Qwen · 2026-10-30
- Learning Zig: explicit allocators as parameters change how you think about memory — arpit_bhayani · 2026-10-30