Anthropic Workbench sunset + Gemini sampling deprecation — act now

🚀 Models & Releases

Anthropic is sunsetting the legacy Workbench (platform.claude.com/workbench) on August 17, 2026. Saved prompts, variables, and evals will not migrate to the new Workbench. Export data via the banner or Organizational Settings before the deadline. (Anthropic API docs)
Google Gemini deprecated temperature, top_p, and top_k sampling parameters. These are now ignored in the latest models. Update inference calls to remove these parameters to avoid silent failures. (Gemini API docs)
Poolside released Laguna S 2.1, a 118B total parameter (8B active) coding agent scoring 70.2% on Terminal-Bench 2.1. Available via OpenRouter. (OpenRouter)
Meituan launched LongCat 2.0, a sparse MoE model with 48B active parameters (1.6T total), optimized for coding, repository-level changes, and agentic workflows. (OpenRouter)
Google released Gemini 3.6 Flash (high-efficiency for coding/agentic workflows) and Gemini 3.5 Flash-Lite (upgraded agentic subagent capabilities). (OpenRouter, OpenRouter)
Microsoft open-sourced Mage-Flow-Edit models (Base, Turbo, full) for instruction-based image editing via rectified flow. MIT-licensed, available on Hugging Face. (Hugging Face)

Evidence-Aware RL paper tackles repetitive copying in long-context reasoning, proposing a method to ground outputs in retrieved evidence. (arXiv:2607.19345v1)
Appearance Pointers enables precise regional control in Diffusion Transformers (DiTs) for multimodal image generation. (arXiv:2607.19344v1)
CodeRescue introduces budget-calibrated recovery routing for coding agents, optimizing cost vs. accuracy in multi-model workflows. (arXiv:2607.19338v1)

Weaviate v1.38.6 includes breaking changes to the REST API response envelope: {id, properties, references, metadata}. Update clients to handle the new structure. (GitHub release)
Anthropic Python SDK v0.117.1 fixes AWS credential handling in AnthropicAWS.copy(). Upgrade if using AWS integrations. (GitHub release)
OpenAI Codex v0.145.0 adds experimental paginated thread history, sub-agent support, and expanded /import for Cursor/Claude Code settings migration. (GitHub release)

EU AI Act: European Commission published guidelines on transparency obligations. Review for compliance if deploying models in the EU. (RAPS)
US: Sen. Rounds introduced the Homestake AI Act to accelerate disease research via AI. (KOTA Territory News)

Anthropic’s Workbench sunset (Aug 17) and Gemini’s sampling parameter deprecation require immediate action to avoid breaking changes in production. Export data and update inference calls now.
LongCat 2.0 (Meituan) and Laguna S 2.1 (Poolside) offer new options for coding/agentic workflows, with LongCat’s 1.6T total parameters standing out for long-horizon tasks.