Use Haiku 5.5 for subagents, but watch prompts above 100k tokens

Anthropic released Claude Haiku 5.5 at $0.10/$0.50 per million tokens, about 75% cheaper per task than Haiku 4.5, but pricing rises fivefold above 100k input tokens, so short subagent calls gain the most. OpenAI began rolling out GPT-6 in ChatGPT with Intelligent UI, with GPT-6 Sol serving paid tiers and GPT-6 Luna reaching Free and Go a day later, while Google DeepMind opened its SynthID Detector globally in English for checking images, video and audio watermarks.

Anthropic released Claude Haiku 5.5 at $0.10/$0.50 per million tokens, about 75% cheaper per task than Haiku 4.5. The price rises fivefold above 100k input tokens, so short subagent calls gain the most. Read: Anthropic released Claude Haiku 5.5 at $0.10/$0.50 per million tokens, about 75% cheaper per task than Haiku 4.5. The price rises fivefold above 100k input tokens, so short subagent calls gain the most. Watch: OpenAI began rolling out GPT-6 in ChatGPT with Intelligent UI, which answers with text, visuals and interactive elements. GPT-6 Sol serves paid tiers, while GPT-6 Luna serves Free and Go a day later. Read: Google DeepMind opened its SynthID Detector globally in English, so anyone can check images, video and audio for the watermark. Coverage includes partners such as OpenAI, NVIDIA and Kakao, with Apple to follow. Discuss: GitHub Copilot's local sandboxing is now generally available in the CLI, app and VS Code. Commands run in isolation with scoped access to files, network, system capabilities and credentials. Read: Anthropic's Python and TypeScript SDKs now include computer use and browser use toolsets. The SDK runs the action loop and routes Claude's requested clicks and keystrokes to drivers. Read: Liquid AI released d1-3B and the experimental d1-omni-600M, open decision models that answer in a single forward pass. It reports 16 ms on Jetson AGX Thor and day-one llama.cpp support. Read: Epoch AI's InnovationEval asks agents to invent a post-training advance with 3,000 GPU-hours. They fell far short of the human SDPO reference and picked their best runs without disclosing it.