Alibaba's new coding-agent benchmark, Co-Design Bench, tests AI agents on real user requests that already tripped them up in practice. Even GPT-6 Astra, the top performer, fully solves only 26.4 percent of tasks. On older benchmarks, the same models score above 99 percent. The article AI coding agents still need human cleanup on nearly three out of four tasks in a real-world test appeared first on The Decoder.
Research1 min read
AI coding agents still need human cleanup on nearly three out of four tasks in a real-world test
Alibaba's new coding-agent benchmark, Co-Design Bench, tests AI agents on real user requests that already tripped them up in practice. Even GPT-6 Astra, the top performer, fully solves only 26.4 percent of tasks. On older benchmarks, the same models score above 99 percent. The…
The DecoderBy: Jonathan KemperNovember 8, 2026 (59 minutes ago)
Available in 4 languagesAI Summary & Key Takeaways
- Alibaba's new coding-agent benchmark, Co-Design Bench, tests AI agents on real user requests that already tripped them up in practice
- Even GPT-6 Astra, the top performer, fully solves only 26.4 percent of tasks
- On older benchmarks, the same models score above 99 percent
Research
KZ·AI 6842Инсайт дня
AI coding agents still need human cleanup on nearly three out of four tasks in a real-world test
From the original source: The Decoder
View originalOpen original
Related stories
AI agents are being tested for their ability to work with real laboratory data
🧬 AI agents are being tested for their ability to work with real laboratory data. In collaboration with biotech company Ligo, a new benchmark BioBasicsBench has been developed, which evaluates how well neural networks handle everyday biological tasks…
Telegram — GPT Time | Нейросети | ИИ
World Models Will Soon Be Inescapable. Here’s What You Need to Know
Moving AI from text to 3D space, world models simulate real-world physics to teach robots and autonomous cars how to navigate reality.
Wired — Artificial Intelligence
Testing OpenAI and Anthropic models on social deduction
📌 Testing OpenAI and Anthropic models on social deduction. Podcaster Dwarkesh Patel spent $20,000 on API to organize a tournament among 11 models playing Blood on the Clocktower, a more complex analogue of Mafia. 🟡Game rules: The whole point is the struggle for information. Participants…
Telegram — AI Machine Learning Big Data
Neural networks are accessible to everyone. Why does business keep paying specialists?
Neural networks are accessible to everyone. Why does business keep paying specialists? Knowing how to use ChatGPT is already a weak specialization. Companies pay for results you can ask about: the program works, leads aren't lost, advertising delivers measurable impact. A beautiful demo…
Telegram — Paragon | Нейросети и ИИ