The Drip
Yesterday was supposed to be the last day Fable Five is included in an Anthropic subscription. That deadline has been moved to July 12th. After that, every Fable call bills by the token, straight to your usage credits or API.
We watched the whole saga play out: it dropped while we were in a workshop with a client, got blocked the next day through a series of government actions, then came back a few days later with tighter guardrails. And now the flagship model is effectively gated behind API pricing.
Justin ran the numbers. Based on his last thirty days of usage, running Fable at that pace would cost somewhere around six or seven grand a month.
Which is the thing we kept circling back to: you pay your subscription and get pretty good results, but if you have real money, you get five times that. The haves-and-have-nots gap is widening. The best tools now have the potential to be very exclusive. That's a strange turn for technology that was supposed to be about empowering everybody.
Inside The Bottle
We just shipped the final 20% of Pathfinder, and it was the most intense build we've done. If there's one myth worth killing, it's that the last stretch is "just bugs." It isn't.
Kellan's reframe: it's not 80/20 anymore, it's three buckets: 80%, then 15%, then 5%. The 80% is the fun part, the one-shot magic where you stand up an enormous amount in a few days and feel like a superhero. The 15% is still hard, but chunkable: there's real meat you can hand off and iterate on. The 5% is the "when is this going to end" part: the pre-go-live grind where, honestly, if you ship without it the thing isn't usable.
The hardest part of that 5% isn't code. It's de-scoping: deciding where the line is. We'd imagined Pathfinder working a dozen different ways, and finishing meant aggressively cutting scope. That creates what we started calling context debt. The models are excellent at execution: wiring up a page, a route, a button. Where they struggle is judgment about what actually matters. One example stuck with us: Claude could not convince itself it didn't need to maintain backward compatibility for users who don't exist yet. If it isn't written down, front and center in the CLAUDE.md file, it won't weigh it.
They also leave fingerprints everywhere. Kellan found 171 memory files the model had created for this one project. When he cleaned them up, 17 were actually needed. We started calling the leftovers Claude crumbs.
The takeaway isn't "AI can't finish." It's that discipline matters more than volume; it's not how much code you ship. Set the rules up front, trust but verify (the same model is genuinely sharper some days than others), and remember that no amount of synthetic testing replaces a real human actually using the thing.
Lab Notes
| ■ | Justin's note: Stop thinking of Claude as a stateful person. It's not "always Claude." It quietly biases toward how each of you likes to work, and when two people's sessions cross-pollinate, it gets confused about what lives where. The unglamorous fix: agree on how you do things (one way to send an email, one place for docs) before you build. |
| ■ | Kellan's note: The same model isn't the same day to day. Opus 4.8 was sharp one morning and brain-dead the next: "that's not even close to what I asked for." Build that inconsistency into your expectations instead of assuming a good result yesterday means a good result today. |
What Stopped Our Scroll
| ■ | Natural Language Autoencoders (Anthropic): a new interpretability method that turns a model's internal activations into plain-English explanations, used in safety audits to catch when a model knows it's being tested or is hiding its motivations. |
|
