Benchmarks

4 articles tagged "Benchmarks"

Mistral AI Launches Codestral 2 — Outperforms GPT-5 on Coding Benchmarks

The French AI company releases a specialized coding model that sets new records on HumanEval, SWE-bench, and real-world software engineering tasks.

Sofia Reyes3h ago2 min read

Research

AI Model Achieves Human-Level Performance on Graduate-Level Mathematics

A new benchmark shows frontier models solving complex proofs and mathematical reasoning problems at the level of a PhD student, with implications for scientific discovery.

Marcus Webb2d ago

Research

Study: Half of AI-Written Code That Passes SWE-Bench Would Be Rejected by Real Developers

Research organization METR has exposed a critical disconnect between benchmark performance and real-world code quality: approximately 50% of AI-generated code that clears the widely-used SWE-bench industry test contains issues that would cause actual project maintainers to reject it, raising fundamental questions about how we measure AI coding capability.

Ryan Torres2h ago

Research

Alibaba's Qwen 3.5-9B Outperforms OpenAI's 120B Model — The Small-Model Revolution Arrives

Alibaba's latest release, Qwen 3.5-9B, is a compact 9-billion-parameter reasoning model that bests OpenAI's gpt-oss-120B on leading reasoning and coding benchmarks while requiring a fraction of the compute. The result underscores a defining trend of early 2026: smaller, hyper-efficient models trained with advanced distillation and reinforcement learning are increasingly matching frontier giants at dramatically lower inference cost. For enterprises, the implications are significant — capable AI is becoming deployable on-premise rather than requiring expensive cloud API calls.

Aria Chen5h ago