AI developer tools · What shipped

As of August 30, 2026, GPT-5.6 Sol leads the BrowseComp benchmark with 92.2%, followed by Kimi K3 (91.2%) and Claude Opus 5 (90.8%). BrowseComp measures web-browsing agent performance on research-oriented questions requiring search, source inspection, and evidence synthesis. The benchmark evaluates 40 models and carries 28% weight within the Agentic category of BenchLM.ai's overall scoring system. Top models are clustered within 1.4 points, indicating the benchmark is approaching saturation for frontier models. BrowseComp 2026 is refreshed quarterly with a public benchmark set.

Claude Opus 5 leads the GDPval-AA benchmark for economically valuable tasks with a normalized score of 66.2%, followed by GLM-5.3 (62.9%) and Qwen3.8-Flash-Next (61.9%). This benchmark, updated August 30, 2026, evaluates 101 models on professional agentic workflows with a quarterly refresh cadence, though it is currently displayed as reference-only and excluded from BenchLM's overall scoring formula.

NVIDIA released several major AI model updates and benchmarks across its Nemotron family. Nemotron 3.5 Lightning is a 30B parameter efficient language model with a 1M-token context window and multi-token prediction layers, optimized for agent deployments and supported across vLLM, SGLang, TRT-LLM, and other frameworks. Nemotron 3 Super (120B total / 12B active parameters) delivers up to 5x higher throughput than the previous version and incorporates LatentMoE and native NVFP4 pretraining. Nemotron 3 Ultra is a 550B frontier-scale model designed for complex multi-agent applications. In speech recognition, Parakeet-tdt-0.6b-v3 now supports 25 European languages with automatic detection, while Parakeet Realtime EOU offers 80–160ms latency for voice AI agent turn-taking with 120M parameters. NVIDIA also released Cosmos 3, a unified omni-model for physical AI with three variants: Cosmos 3 Super (32B reasoner + 32B generator), Cosmos 3 Nano (8B class), and Cosmos 3 Edge (4B class) for real-time robotic deployment. Cosmos pairs an autoregressive reasoner with a diffusion generator in a single forward pass across text, image, video, and action prediction. Additionally, NVIDIA published comprehensive open datasets including Nemotron-Math-v2 for reasoning, Nemotron-SFT-Code-v3 for code generation, and HelpSteer3 for reward modeling, alongside evaluation benchmarks like SPEED-bench for model performance assessment.