BTC
ETH
SOL
XRP
BNB
DOGE
ADA
AVAX
LINK
TON
DOT
LTC
TRX
POL
USDT
USDC
SUI
HBAR
SHIB
BCH
XLM
NEAR
APT
ARBA fixed set of tasks used to test and score AI models, so different models can be compared on the same work.
AI Drafted with AI help · Reviewed by Keshav Aggarwal · Updated
In simple words
Think of a school exam that every student in the country takes on the same day. Because the questions are the same for everyone, you can compare who did better.
An AI benchmark is that exam for AI models. But if the same exam is used for years, students can learn the answers, so new, harder exams are needed.
The technical side
A benchmark is a fixed set of tasks with a scoring rule, used to compare AI models on maths, coding, reasoning, safety or agent work. Labs quote benchmark scores when they release a model. When top models all score close to 100%, a benchmark is called “saturated” and stops telling them apart, so researchers build harder ones; a model can also do well only because test questions leaked into its training data.
On 8 November 2026 The Decoder reported on Co-Design Bench, built by researchers from Alibaba and several Asian universities, which asks coding agents to change existing web apps from real user requests. An AI judge tests each finished app in a browser and a second judge re-checks every pass. The best of ten models, OpenAI’s GPT-6 Astra, met about 70% of a task’s requirements on average but fully completed only 26.4% of tasks, while models score above 99% on older coding tests. Compare models on AI models.
AI Benchmark kya hai?
AI benchmark ek fixed exam hai jisse alag-alag AI models ko same sawaalon par test karke number diye jaate hain. Jab sab models purane exam mein 99% laane lagte hain, to naye aur mushkil benchmark banaye jaate hain, jaise Co-Design Bench, jismein sabse achha model sirf 26.4% kaam poora kar paaya.
Related terms
- LLM (Large Language Model)AIAn AI model trained on huge amounts of text so it can read and write language, like the models behind ChatGPT, Claude and Gemini.
- AI AgentAIAn AI system that can plan steps and use tools on its own, such as searching, clicking or sending messages, to finish a task.
- Reasoning ModelAIAn AI model that works through a problem step by step before answering, which helps with maths, code and planning but takes longer.
- HallucinationAIWhen an AI model states something false or made up as if it were true, such as a fake fact, quote or source.
- Red TeamingAITesting an AI by attacking it on purpose, trying to make it break rules or cause harm, so the problems are fixed before the public uses it.