After 5 weeks and 3 postponed deadlines, Anthropic finally landed on an answer for Fable 5 access.

Max and Team Premium subscribers keep Fable at half their plan's usage caps. Lower plan users get a one-time $100 credit before moving to pay-per-use.

Anthropic called the demand "challenging to predict" and admitted the situation was "frustrating." Honest, at least.

Sam Altman had a field day on X. Posts like "clarity is nice" and "stay because we don't treat you with contempt" landed while Anthropic kept pushing the cutoff date. Good timing on his part.

The model is 6 weeks old. Kimi K3 is at the frontier now. GPT-5.6 Sol's usage limits keep expanding. Anthropic is building more compute, but the access question will probably follow the next launch too.

The bigger signal: 👉 Compute capacity is the bottleneck nobody talks about enough. Building the best model means nothing if you can't serve it.

Top engineers at Anthropic and OpenAI say AI now writes 100% of their code.

If you're not using AI, you're spending 40 hours doing what they do in 4.

These 100+ Claude Code hacks fix that and help you ship 10x faster.

Sign up for The Code and get:

Claim your free playbook

Most people burning through their Claude Code budget aren't using the wrong model. They're feeding it too much noise.

Here's the free open-source stack that actually fixes it:

Start with measurement

You can't cut what you can't see. ccusage parses your Claude Code session logs and breaks down usage by model, project, and billing block. Run it first. Then run it after every change.

Install: npx ccusage@latest

Cut shell output noise

RTK (Rust Token Killer) sits between your terminal and Claude, stripping redundant output before it hits the model. Git status drops around 80%. Test runners drop around 90%. Full sessions average around 80% reduction.

Install: brew install rtk then rtk init -g

Make Claude reply like it means it

The Caveman skill forces terse, high-signal replies. No filler, no hedging, no padding. Technical facts and code stay intact. Average output token reduction: around 65%. And because outputs become future inputs, the savings compound.

Many power users leave this on permanently.

Stop Claude from re-reading your whole codebase

Graphify builds a queryable knowledge graph of your project. Instead of reading files from scratch each time, Claude queries the graph. Up to 71x fewer tokens per query on large codebases.

Install:

Add the Karpathy principles

This CLAUDE.md drops a short set of principles into every project: think before coding, make surgical changes, stay goal-driven. Indirect savings, but they're massive. Fewer wrong turns means fewer correction loops.

Drop it into every project you work in.

For long sessions: pxpipe

pxpipe converts heavy parts of your request into dense PNG images. Claude reads them with vision instead of text tokens. Real-world bill reductions of 59 to 70% on production traces.

30-second setup:

One honest caveat: vision is lossy for exact strings like hashes and IDs. Keep critical verbatim data in recent text turns.

The stack most people should start with:

Beyond the tools, two habits matter more than any repo:

End every meaningful session with a compact handoff note covering what changed, what's unverified, and what to avoid next time. And use /clear aggressively when switching tasks. A fresh context costs less than a polluted one.

Still paying GoDaddy or Namecheap prices? Porkbun sells most domains at cost for low, transparent registration and renewal pricing with no nonsense. Get free features like WHOIS privacy and SSL certificates, plus real human support 24/7, 365 days a year. Save $1 on your next domain name now.

Get Your Next Domain Now

The Jacobian Conjecture has been open since 1939. A Harvard-trained mathematician working at Anthropic just posted a counterexample, with Claude Fable 5 as a direct collaborator in the work.

Multiple prominent mathematicians verified the result, including Stanford's Jared Duker Lichtman. Read the full breakdown here.

No hype needed on this one. An 87-year-old open problem in mathematics, closed with AI help. That's just a good result.

The bigger signal: 👉 The most interesting AI contributions right now aren't autonomous. They're human-AI collaborations where the human brings the domain depth and the AI handles the computational weight.

A few tools quietly worth exploring:

Fable access is settled, finally.
An 87-year-old math problem got closed with AI help.
And if you're hitting token limits, there's a free stack that cuts 60 to 90% of the waste.

The tools keep getting better. The cost of using them badly keeps going up too.