A simple prompt file called the "reflection engine" is making rounds in AI circles, and the results are unsettling in the best possible way. The concept is straightforward: download the markdown prompt, upload it to your AI agent, and ask it to evaluate and complete all tasks. What comes back is a surprisingly detailed psychological profile synthesized from everything the agent has access to — therapy transcripts, saved notes, conversation logs, whatever you've fed it.

The Setup and What It Surfaces

The experience was tested across two agents: Fable High and Sol Max. Sol produced the more coherent output — clearer connections, better structure — while Fable's report was harder to parse. This tracks with a broader pattern: as agents grow more capable, output quality varies widely by model and task type.

What made the reflection engine compelling wasn't just novelty. With access to therapy transcripts and memory files accumulated over time, the agent synthesized patterns the user hadn't consciously connected. The result: a 40+ question follow-up session designed to address the findings and build an action plan.

This is a genuinely different use case than task automation. It's using an AI agent as an introspective tool — less "do this for me" and more "tell me what you see in what I've already told you."

The Model Economics Shift

Separately, OpenAI cut prices significantly on its latest model tier. GPT-5.6 Luna dropped 80% in price; Terra came down 20%. Luna at max thinking effort now performs roughly on par with GPT-5.4 xhigh — which was the best available model just four months ago — but at only 8% of the cost.

That math matters for anyone running agents at scale:

  • 10–12x more compute for the same budget if you lean into Luna for non-building tasks
  • Luna Max performs well for research, reading/writing, and general chat
  • Complex coding tasks still benefit from Sol High — Luna made a mess of a Chrome extension that Sol had to clean up

OpenAI also teased a new major model called Astra, releasing solutions to 10 long-standing problems in mathematics and theoretical computer science as a preview of its capabilities.

New Models Worth Tracking

The broader model landscape is moving fast:

  • DeepSeek V4 Flash: $0.14/$0.28 per million input/output tokens, 1M-token context window — undercuts Luna on price
  • Qwen3.8-Max: A 2.4 trillion parameter model claiming near-frontier performance for coding at $2/$6 per million tokens; open weights for smaller variants coming next week
  • Inkling-Small: Open-weight, activates 12B of 276B parameters, reportedly matches the larger Inkling on many benchmarks
  • MiniMax H3: New video model positioned against Seedance, with open weights promised soon
  • P-Image-Ideogram: Four quality modes, native 1K/2K output, starting at $0.003 per image

What This Means for Builders

Google launched Gemini Robotics 2 — a single AI system designed to run across robot arms and full humanoids, with one model for translating vision and instructions into movement, another for complex task planning, and an on-device variant for edge deployment.

On the tooling side, several launches are worth noting for teams building with agents:

  • YC open-sourced a multi-agent harness it uses across accounting, legal, events, and engineering
  • Conductor Cloud offers multiplayer cloud workspaces for Claude Code, Codex, Cursor, and OpenCode that persist after you close your laptop
  • Comp AI's CRM is an open-source, agent-first CRM with durable research agents under MIT license
  • Cloudflare Computer is an open-source virtual filesystem and execution runtime for agents, built on Workers

The reflection engine experiment, priced-down frontier models, and a wave of open-source tooling all point in the same direction: the cost of running sophisticated agents is collapsing, and the surface area of what you can do with them is expanding well beyond code and content into territory that feels genuinely personal.