New AI Agent Shocked The Industry: Crushed GPT5 Codex and Claude

Abacus AI's new DeepAgent Desktop has surpassed major AI models like GPT-5 Codex and Claude code in coding benchmarks and offers a comprehensive suite for software development.

• DeepAgent Desktop from Abacus AI currently leads AI coding benchmarks.

• The system combines CLI, code editor, and chat modes with an integrated testing agent.

• It offers a comprehensive, flexible, and affordable AI development environment.

Abacus AI has launched DeepAgent Desktop, an AI coding agent that has outperformed GPT-5 Codex and Claude code on Terminal Bench and SWE Bench Verified, achieving the highest scores to date. DeepAgent Desktop is a full desktop suite combining CLI, code editor, and chat modes, notably featuring its own testing agent for code validation. This innovative approach integrates multiple AI models and provides an all-in-one development environment, setting a new standard for AI-assisted software engineering.

DeepAgent Performance

• 00:00:01 DeepAgent Desktop, developed by Abacus AI, has significantly outscored leading AI models such as GPT-5 Codex and Claude code in critical benchmarks. It achieved 48.75% on Terminal Bench, surpassing GPT-5 Codex's 42.8% and Claude code Opus's 43.2%. On SWE Bench Verified, the standard for automated bug fixing with real-world GitHub issues, DeepAgent scored an impressive 74%, exceeding GPT-5 Codex's 72.8% and Claude Code Sonnet 4's 72.7%.

Integrated System Modes

• 00:01:42 DeepAgent Desktop is not a single model but a comprehensive desktop suite integrating three distinct modes: CLI Agent, a code editor agent, and a chat mode. A key differentiator is its included testing agent, which autonomously tests generated code to ensure functionality. This integrated approach, which allows access to external models like Claude, Gemini, and GPT-5, significantly contributes to its strong benchmark performance and overall utility.

CLI and Code Editor Capabilities

• 00:02:14 The CLI mode is branded as the fastest way to code, demonstrated by its ability to generate complex projects like a snake web game with a retro Nintendo vibe or a full LinkedIn clone ('Connect Hub') from single prompts. The code editor mode functions as an AI-powered IDE, capable of tasks such as creating a personal website from an uploaded resume image, involving OCR and data extraction, or generating an advanced holistic guide for 'vibe coders' as an interactive website with animations and resources.

Chat Mode and Accessibility

• 00:05:17 The chat mode transforms DeepAgent Desktop into an all-in-one AI assistant, offering flexibility by allowing interaction with DeepAgent's internal system alongside external models like Claude, Gemini, and GPT-5 within a single interface. Demonstrations include generating a detailed 90-day playbook for LinkedIn and Twitter influence in software development, providing step-by-step instructions for setting up open-source game repos, and explaining complex topics like spacecraft landing on Mars with rich detail. Abacus AI has made DeepAgent Desktop accessible, pricing the basic tier at $10 per month and fostering community engagement through weekly competitions with cash prizes.