AI developer tools · What shipped

Added Claude Opus 5 (claude-opus-5) as the new default Opus model with 1M context and fast mode pricing at $10/$50 per Mtok. The release notes cover extensive bug fixes and improvements across Claude Code v2.1.223 through v2.1.214, including security fixes for permission bypasses in Bash and PowerShell, sandbox isolation improvements, MCP server enhancements, and better handling of workflow agents and background sessions. Notable infrastructure updates include improved permission checking for parallel tool calls with better cache efficiency, enhanced session management for forked and background agents, and refined auto-mode safety with the permission classifier now evaluating messages sent between agent sessions before dispatch.

Claude Opus 5 launched July 24, 2026 at $5/$25 per million tokens (input/output), matching Opus 4.8 pricing and leading Frontier-Bench v0.1 coding evaluations according to Anthropic's benchmarks. Claude Sonnet 5 released June 30, 2026 at $3/$15 standard ($2/$10 intro through August 31) as the most agentic Sonnet model yet, now default on Claude Free and Pro plans. OpenAI cut GPT-5.6 pricing by 80% for Luna to $0.20/$1.20 and 20% for Terra to $2/$12 on July 30, framing the cuts as efficiency gains passed through to users. GitHub shipped customizable reasoning levels for Copilot's cloud agent (August 3), comment-triggered automations from issue/PR comments (August 3), and moved agent skills with MCP server support for code review to general availability (July 29). Google added environment hooks for agent sandboxes, token-spend budget caps, scheduled cron triggers, and free-tier support to Gemini API Managed Agents (July 28). Separately, GitHub retired Copilot Billing Preview and deprecated GitHub Spark (August 4), while Gemini 2.5 Pro and Gemini 3 Flash were deprecated as Copilot model options (July 31).

Tencent announced global availability of Hy3, a 295-billion-parameter AI model with 21 billion active parameters built on a hybrid Mixture-of-Experts architecture supporting 256K context length. The model performs comparably to flagship models with two to five times more parameters across reasoning, code generation, and agent capabilities. Hy3 is accessible via API through OpenRouter starting at $0.1288 per million input tokens, integrated into products including WorkBuddy (free until August 31, 2026), Tencent Design Miora, and Tencent Cloud TokenHub. The model achieved first place on OpenRouter's global LLM usage leaderboard within one week of launch and recorded 68 times more API calls than its predecessor. Developers can access Hy3 through Hugging Face, ModelScope, and third-party platforms including Hermes, Kilo, and Cline under an Apache 2.0 license.

Cloudflare announced Cloudflare OS, an open-source AI workspace available now on GitHub that runs on Cloudflare's global network. The platform provides employees with secure AI tools and access to internal company systems without requiring new infrastructure, featuring an Agent Workspace accessible via browser, app-building capabilities without developer expertise, Zero Trust security via Cloudflare Access, and multi-model AI support with cost controls through Cloudflare AI Gateway. Cloudflare OS was built internally for Cloudflare's workforce and is now offered as open-source software with managed deployment options coming soon, supported by implementation partners including Presidio and Happy Cog for enterprise customization.