The Drip
Fable 5 has officially left the subscription. Anthropic's deadline for using Fable as part of a regular plan was supposed to be July 7. It got a last-minute extension to July 12. Oh, and wait, it was extended to July 19….
Either way, the era of "best model included" is likely over, and the new practitioner skill is understanding the usage-credit puzzle. Our short version: if you're not hitting limits every week, consider a cheaper plan plus usage credits, and spend Fable where it's strongest, planning and reviewing, while cheaper models do the implementing. We get into the details on this week's episode.
Inside The Bottle
Pathfinder is real. It's live, it's free for individuals, and this week we finally take you on the full tour.
The backstory, in one line: after a year of AI conversations that all sounded like "we have Copilot or ChatGPT, but we're still figuring out what we'd use it for," we became convinced most teams don't have an AI problem. They have a focus problem. People aren't stuck because the technology fails them. They're stuck in evaluating, thinking, wondering. Pathfinder is our system for getting from that state to acting.
It walks you through the five states of the work: Idea, Focus, Build, Change, Review.
Idea. The brainstorm flips the usual setup: the AI interviews you. One question at a time about your role and where the week actually goes. Give it 15 minutes and you'll leave with six to ten real ideas. No blank page, no homework.
Focus. Before anything gets built, one idea at a time gets drilled into a Brief: the real problem, what solving it would mean for you, and what kind of solution actually fits. Not everything needs AI in the day-to-day. Some problems want a plain automation, some an assistant, some an agent, and some just need an app that makes the problem go away. Not five ideas. One.
Build. The Build Plan expands your Brief for whatever tool you build with: hand it to Claude Code, Codex, or ChatGPT by copy-paste, MCP connector, or CLI. The push is always toward the smallest first version that's still valuable. Rich, focused context for the session where the real building happens.
Change and Review. A roadmap sets a reasonably ambitious 90-day destination, and checkpoints keep you honest: did you do the thing you committed to? Review, recast, or close it out.
Individuals ride free. That's deliberate. The team plan adds synthesis and prioritization across your whole group, plus us, when you want humans who've been through the wringer involved.
Try it: pathfinder.tinybottleai.com
Lab Notes
| ■ | Justin's note: Fable wrote itself a memory that it always needed my permission before changing our system, all from one moment where I said "hold on." A moment isn't a rule. Worth checking what Claude writes down about you. |
| ■ | Kellan's note: Leave the extra-high and ultracode effort settings alone. They'll token-max their way to the ends of the earth, and the benchmarks say high (sometimes medium) actually performs better. Human in the loop stays part of the conversation. |
What Stopped Our Scroll
| ■ | OpenAI shipped ChatGPT Work: pulls context from Slack, Teams, Drive, and CRMs, and independently produces docs, slides, and sites over hours, not minutes. Atlas browser killed to make room for it. |
| ■ | Meta shipped Muse Spark 1.1, its first "agentic era" model: multi-agent orchestration, 1M context. |
|
