AI developer tools · What shipped

MCP Atlas is a benchmark measuring tool-calling capabilities of AI models over Model Context Protocol integrations and external tools. As of September 10, 2026, 41 models have been evaluated, with Muse Spark 1.1 leading at 88.1%, followed by Claude Opus 5 (85.8%) and Kimi K3 (84.2%). The benchmark falls under the Agentic category and is currently displayed for reference but excluded from BenchLM.ai's overall scoring formula.

SRE Bench, a new contamination-free reverse engineering benchmark from Vals AI, evaluates AI agents' ability to determine what real-world binaries do without source code access. The benchmark comprises 262 instances with 1,572 tasks across network protocols, firmware, games, file-format recovery, and malware, built from 19 in-house programs averaging 16,915 lines of code and protected with state-of-the-art anti-analysis techniques. Results show GPT-6 Astra leading at 56.87% accuracy ($13.50 per instance), followed by GPT-5.6 Sol at 30.53% and Claude Fable 5.1 at 22.90%, with performance degrading significantly on protected binaries and malware samples, revealing current limitations in agentic reverse engineering capabilities.

MATLAB Agentic Toolkit released version MATK-2026.09.a with new skills including matlab-write-tests for generating unit tests using class-based frameworks and matlab-run-tests for running test suites and collecting code coverage. The AI and Statistics group added matlab-analyze-reliability for fitting life distributions. Previous release 2026.08.27 introduced matlab-engineer-tabular-features for feature engineering and matlab-use-symbolic-math for generating MATLAB code using Symbolic Math Toolbox. Release 2026.08.13 added matlab-analyze-spectral-images for processing hyperspectral and multispectral images, and consolidated 22 skills with improved naming consistency. The 2026.07.31 release expanded the toolkit with new skill groups for control systems and environment settings, plus new AI/Statistics skills including matlab-classify-tabular-data for comparing candidate models and matlab-import-external-ai-model for importing PyTorch, ONNX, and Keras models into MATLAB. Code generation improvements include matlab-generate-code for C/C++/CUDA code from MATLAB and matlab-deploy-ai-model for generating code from PyTorch models. Earlier releases in June and May 2026 introduced the Agentic Toolkit Installer reducing context window usage, RF and Mixed Signal plugin expansions with 26 new skills, and Automotive skills for RoadRunner HD map authoring and scene import.

Claude Fable 5.1 and Claude Mythos 5.1 launched in September 2026 as the world's most advanced models for coding and knowledge work. Claude Sonnet 5 launched in June 2026 with substantial improvements in reasoning, tool use, coding, and knowledge work. Claude Opus 5 launched in July 2026 as a thoughtful and proactive model approaching frontier intelligence at half the price of Claude Fable 5. Claude Code now supports computer use capabilities, allowing it to open files, run dev tools, and navigate screens to perform tasks. The Code Execution Tool launched in August 2025, enabling Claude to execute Python code in a secure, sandboxed environment. Claude Cowork became generally available on macOS and Windows with role-based access controls for Enterprise plans and expanded to web and mobile in July 2026.