AI developer tools · What shipped

Claude Opus 5 has been released and tested in a comprehensive benchmark covering prototype creation, PRD development, wireframe design, bug triage, and agentic coding tasks. In the "How I AI Benchmark," Opus 5 ranked at the top of the leaderboard, outperforming GPT-5.6, Sonnet 5, and other frontier models, particularly excelling at front-end design and app prototyping work. However, the model exhibits notably cautious and apologetic behavior compared to GPT models, frequently delegating decisions back to users and showing high human dependency. The review includes a 70-30 split between human evaluation and AI-based judging, with Opus 5 scoring highest despite user frustration with verbose output and excessive hedging in responses.

AWS announces a research preview of aws-bench, an open-source benchmark for measuring AI agent performance on AWS tasks. The tool provides a suite of test cases derived from real AWS usage patterns, pairing natural-language queries with defined cloud resource states and ground-truth answers to enable objective, reproducible evaluation of models and agents. The release includes a CLI tool for instantiating testing environments, executing evaluation runs, and scoring agent performance, and is available now on GitHub.

Jupyter AI v3.1.0 released with enhanced in-chat controls for collaborative environments. AI personas now expose all settings (model, permission mode, etc.) in the input toolbar and respond reliably without requiring @-mentions. The release adds support for JupyterLab v4.6 and compatibility with Voilà dashboards, plus major bug fixes to the underlying jupyter-server-documents RTC engine. Earlier, Jupyter AI v3.0.0 introduced agent support via the Agent Client Protocol (ACP), enabling Claude, Codex, Gemini, Goose, Kiro, and OpenCode in JupyterLab. The v3.0 release featured real-time chat UI with live-streaming responses, tool call permissions, a new Jupyter MCP server (jupyter_server_mcp) for agent integration, and support for custom MCP servers via .jupyter/mcp_settings.json configuration.

Nsight Systems 2026.4.1 has been released with enhancements to GPU performance analysis tools. New features include D3D12 interfaces support (ID3D12Device15, ID3D12ApplicationIdentity, ID3D12Runtime Validation Control), Vulkan extensions (VK_KHR_maintenance10, VK_EXT_memory_decompression, VK_NV_compute_occupancy_priority, VK_NV_low_latency2), and improved CUDA stream visualization with NVTX projection on combined stream hierarchies and cuTile name demangling. PyTorch profiling gains a new "–pytorch=function-trace-shapes" option for detailed function shape tracing, while network improvements include high-frequency NIC metrics collection and NCCL straggler detection. An Agentic AI skill pack preview focuses on graphics performance and frame stutter analysis designed to work with leading AI agents and models.