ARTFEED — Contemporary Art Intelligence

AI Agents Compile Behavioral Records: Kevin Rose's Reflection Engine and OpenAI's GPT-5.6 Price Cuts

ai-technology · 2026-08-04

Kevin Rose has created a prompt called 'Reflection Engine' that compiles a complete behavioral record of a user from their AI agent's data. The prompt, available as a markdown file, instructs the agent to evaluate the file and complete tasks, resulting in a report based on the user's files, including therapy transcripts and other 'memories'. Rose tested it with Fable High and Sol Max, finding Sol's report more coherent. He shared parts of his own report and recommends others try it. Meanwhile, OpenAI reduced the price of GPT-5.6 Luna by 80% and Terra by 20%, with Luna at max thinking effort scoring similarly to GPT-5.4 xhigh at 8% of the cost. OpenAI also teased Astra, a new model that solved 10 long-standing problems in maths and theoretical computer science. Google launched Gemini Robotics 2, an AI system for controlling robots. Other model releases include DeepSeek V4 Flash, Qwen3.8-Max, Inkling-Small, MiniMax H3, and P-Image-Ideogram. The newsletter also mentions various tools and projects, including a marketing team template for agents by Vercel, YC's open-source multi-agent harness, and ChatGPT desktop app's new sidebar view.

Key facts

  • Kevin Rose created 'Reflection Engine', a prompt that compiles a behavioral record from AI agent data.
  • The prompt is a large markdown file that users upload to their agent and ask to evaluate.
  • Rose tested the prompt with Fable High and Sol Max; Sol produced a more coherent report.
  • OpenAI reduced GPT-5.6 Luna price by 80% and Terra by 20%.
  • Luna at max thinking effort scores similar to GPT-5.4 xhigh at 8% the cost.
  • OpenAI teased Astra, a new model that solved 10 long-standing problems in maths and theoretical computer science.
  • Google launched Gemini Robotics 2, an AI system for controlling robots.
  • DeepSeek V4 Flash costs $0.14/$0.28 per million tokens with a 1M-token context window.
  • Qwen3.8-Max is a 2.4T model with weights coming next week.
  • Inkling-Small is open-weight, activating 12B of 276B parameters.
  • MiniMax H3 is a new video model matching Seedance's features for less.
  • P-Image-Ideogram offers four image-generation quality modes starting at $0.003 per image.
  • Vercel built a marketing team template for agents using Eve framework.
  • YC open-sourced a customisable multi-agent harness for accounting, legal, events, and engineering.
  • ChatGPT desktop app has a new sidebar view for conversations needing attention.

Entities

Artists

  • Kevin Rose

Institutions

  • OpenAI
  • Google
  • DeepSeek
  • Qwen
  • Thinking Machines
  • MiniMax
  • Krea
  • Ideogram
  • Vercel
  • Y Combinator
  • Supabase
  • Cloudflare
  • Comp AI
  • FutureSearch
  • Monologue
  • GenOffice
  • Conductor Cloud
  • Nutrient

Sources