My thoughts on Fable

Fable, Anthropic's brainstorming companion, has one more free day riding inside Claude subscriptions before it switches to paid usage credits. Ben's Bites writer Ben Tossell spent the week leaning on it not to build anything, but to think out loud — closer to therapy session than coding sprint. He noticed something odd: Fable started showing Opus-like reasoning patterns even though he confirmed his prompts weren't being routed to that model. Whether that's a genuine shift in behavior or just pattern-matching from someone staring at chat logs too long is unclear. But it's the kind of detail that makes you wonder how much invisible tuning happens behind the scenes of these tools between releases.

The bigger story running underneath is the search for a harness that thinks rather than just codes. Tossell tried stripping down Claude Code's system prompt to make it faster and less developer-brained, hoping for something that could spar with him creatively instead of just executing tasks. He compared Codex favorably to Claude Code on raw speed and polish, yet kept circling back to the same problem: coding agents have coding brains, and forcing one into creative brainstorming is like handing a carpenter a paintbrush. Plain Claude didn't solve it either. Skills, custom prompts, different harnesses — none of it clicked.

Meanwhile, GPT-5.6 is expected to land this week, possibly as GPT-5.6-Sol, and the newsletter frames it as another shot at cracking the same creativity puzzle. Anthropic also dropped new interpretability research showing Claude activates a kind of global workspace — clusters of concepts that never surface in the visible output or chain-of-thought but still shape its answers, something the researchers likened to unconscious thought. OpenAI, separately, reportedly halved its inference costs and is already testing the change on logged-out ChatGPT users, per The Information. The last time OpenAI cut prices this aggressively was a year ago, an 80% drop on o3.

The smaller items paint a picture of an ecosystem racing to make agents feel less like tools and more like collaborators: a menu bar app tracking Claude and Codex spend, a new realtime OpenAI model that reasons over video and outputs audio, a Rocket League-style playable world model trained on 10,000 hours of gameplay, and a growing catalogue of MCP integrations meant to give agents more hands to work with. None of it answers Tossell's core question — what do you actually want when you sit down to create something — but it does suggest a lot of people are betting the answer will just emerge if you keep cycling through enough AI-generated options.