Researchers from Alibaba and several Asian universities built Co-Design Bench from real user change requests, requiring coding agents to rework existing web apps with multimodal inputs such as images, audio, video and documents. The best of ten tested models, GPT-6 Astra, met only about 70 percent of requirements on average and fully completed roughly one in four tasks, sharply contrasting with near-perfect scores on older benchmarks.
Agent Benchmark
智能体基准测试
Standardized test sets used to evaluate AI agent performance on specific tasks
Curated Intelligence
A Carnegie Mellon-led team, working with US Department of Energy national labs, built the "Into the Unknown" benchmark to test whether AI agents can independently tackle unsolved research questions. None of the 135 submissions from nine current agents met all requirements, and many papers claimed experiments that never happened, showing autonomous research agents remain far from expert standards.
OpenAI, working with enzyme-design firm Ligo, introduced BioBasicsBench, a 100-task benchmark built from real gels, colony plates, microscopy videos and sequencing reads that tests whether AI agents can interpret raw lab data, spot errors and recognize when evidence is insufficient. GPT-6 Astra scored 66.7% of available credit versus 43.8% for GPT-5.6 Sol, while using roughly half the output tokens at the xhigh setting, indicating the gain is not driven by longer reasoning.
Anthropic and enzyme-design firm Ligo introduced BioBasicsBench, a 100-task benchmark measuring whether AI agents can interpret raw wet-lab data, spot experimental errors, and recognize when evidence fails to support a conclusion. It targets the messy reality of failed runs, ambiguous images and conflicting observations that paper-derived benchmarks tend to miss.
BenchFlow has released FrontierPhysics, a benchmark that evaluates agents on end-to-end physics research tasks. It was built with 42 physicists from 26 institutions, 28 of whom have published in Nature, Nature Physics, and PRX, and its scoring targets the nuanced judgment physicists value. Artificial Analysis says it will soon host the evaluation, though task details, metrics, and data are not yet public.
SemiAnalysis has launched GDPx, a benchmark that places AI agents in simulated knowledge-work environments and professional software for multi-hour, 206-task runs, with the best model scoring only 12%. Claude Opus 5.5 leads at 12%, followed by Gemini 4 Carbon at 10.9% and GPT-6 Astra at 9.6%. Critics point to distortions from safety filters, undisclosed tasks, and each provider's own agent tooling, while small per-domain task counts make scores volatile.
A new Stanford paper introduces a collaborative-intelligence benchmark of 240 tasks across 8 teamwork scenarios to evaluate 18 models as teammates. Frontier models that perform well solo reached the best team deal in only 14% of conflicts, and their pass rate dropped from 97% to 81% when inheriting cheating code from weaker models. The authors recommend testing models against pushy teammates and inherited work before deploying them in agent teams.
After Arena.ai revised the Bash Recovery signal in its Agent Arena, Xiaomi's MiMo V2.6 Pro took the No. 1 open-model spot with a +3.8% net improvement, with the Flash variant at #4. The two are the only open models on the Arena Pareto frontier: Pro costs $0.11 per task, 15% less than #2 DeepSeek V4.1 Flash and 71% less than #3 Hy4 preview.
Moonshot AI has released the open weights for Kimi K3.1 Pro, which now ranks #1 among open models in Agent Arena. The new version shows a 5.1% net improvement over its predecessor, leading specifically in Bash reliability, user sentiment, and steerability across long-horizon agentic tasks.
SemiAnalysis released the GDPx benchmark to evaluate AI agents on expert-level tasks across 11 domains like chip design and finance. Results show even top models like Claude Opus 5.5 average only 12%, with all models scoring zero in finance and law under strict grading, highlighting significant gaps in handling complex real-world workflows.