The model is splitting into tiers

Another big day of releases.

OpenAI's Intelligent UI: GPT-6 renders components instead of writing paragraphs. Ask about a 7-speed bike and you get a diagram you can tap through. Wow!

Haiku 5.5: 90% cheaper at $0.10 in, $0.50 out. A year ago Haiku scored 0% on Terminal-Bench. Today it beats GPT-6 Luna on computer use and is built to run as a subagent under Opus or Fable. Another wow! :-)

OpenRouter's decision leaderboard: Jev at 982M requests while #2 is at 12.4M. OpenAI's Decisions API is available now too. I've only tried Jev so far but hearing good things about Cloudflare's Clef.

Just like monoliths did, the model is splitting into tiers:

Presentation, Worker ($0.10 in, $0.50 out), Decision ($0.10, input only) and Planner (Opus/Fable rates).

The decision tier is where its still interesting.

If you've built with Claude Code, you've probably used hooks like PreToolUse or UserPromptSubmit. Most of them run static rules: allowlists, path checks, or a human in the loop. But they see the shape of an action, not its meaning.

I wonder if a decision model could sit behind that hook and score the action. Allow: 0.97. Deny: 0.03.

The hook enforces. The model scores. And the policy sets the threshold.

I gotta try it this week.

The agent is seemingly becoming a system of models.