AI developer tools · What shipped

VIBE-Pro is a repo-level code generation and full-project delivery benchmark that spans web, mobile, and simulation-style implementation tasks. As of August 4, 2026, MiniMax M2.7 leads the public snapshot with a score of 55.6%, with 24 confirmed model releases in the last 30 days. The benchmark measures whether models can complete substantial product requirements across end-to-end software delivery scenarios rather than single-file code snippets.

Flame-VLM-Code is a vision-language coding benchmark for generating correct code from visual and multimodal inputs, with results tracked as of August 4, 2026. Claude Opus 4.6 leads the benchmark at 98.8%, followed by GLM-5V-Turbo at 93.8% and Kimi K2.5 at 88.8%. The benchmark currently covers 3 evaluated models and falls within the multimodal and grounded category, though BenchLM displays it as a reference-only benchmark pending full verification of source attachments.

Claude Code v2.1.222 was released with multiple bug fixes and reliability improvements focused on isolation, permissions, and connectivity. Key changes include fixes for worktree-isolated sessions preventing destructive git commands, PreToolUse auto-allow hooks bypassing tool restrictions, usage credit request blocking on Team/Enterprise, HTTPS proxy connectivity issues, and connection mid-response errors. The release also addressed MCP server usage attribution, pull request linking after branch pushes, org-restricted model handling, stream idle timeouts on custom gateways, and issues with connector authorization, tool error display, long summaries, and file watcher crashes. Additional improvements cover permission classifier safety for cross-session messaging, sandbox credentials masking on Linux/WSL, and various UI fixes for screen readers and diff views. Previous v2.1.221 added Focus view for chat activity summaries, sandbox credential file masking, and MCP server connection improvements in print mode, alongside fixes for Bash permission check bypasses, PowerShell path handling, thinking toggle persistence, and team spend-limit messaging. v2.1.219 introduced Claude Opus 5 as the default model with 1M context and fast-mode pricing, plus nested subagent forwarding and new safety features. v2.1.212 changed /fork to create background sessions and added /subtask for in-session subagents, session-wide WebSearch limits, automatic background promotion for long-running MCP calls, and session resume picker functionality.

Sinch announced Agent Tools, a new developer toolkit for building applications with Sinch's communication platform. The suite includes extensions for Visual Studio Code, JetBrains IDEs, and Open VSX-compatible editors, with built-in support for AI coding assistants including Claude Code, Cursor, GitHub Copilot, and ChatGPT Desktop. Key features include Simulator Mode for testing integrations without creating an account or making live API calls, MCP server integration providing AI assistants access to live Sinch API definitions, and structured product knowledge through Sinch Skills covering configuration patterns across services including Conversation API, Voice, Verification, Numbers, and others.