▶ YouTube Summaries

• https://www.youtube.com/watch?v=3sTu8sSLVfg
• Original title: Mythos is here, it’s time to start tokenmaxxing

Theo's guide to extracting maximum value from heavily subsidized Claude Code subscriptions before Fable (aka Mythos) is removed from the Pro/Max tiers on June 23rd. He frames it less as cost-efficiency and more as a 10-day window to taste what cost-no-object agentic development feels like.

The setup

Fable/Mythos is in Pro/Max Claude Code subscriptions only until June 22 (removed June 23) — too compute-heavy to keep subsidizing. Theo claims a $200 plan can yield up to ~$8,000/month of inference; he burned $4,358 on his laptop plus more on a Mac Mini in 10 days. Disclaimer: nobody should spend $10k/mo personally long-term, but $200–400 for that much inference is compelling during the window.

Gaming the rate limits

• Two limits: a 5-hour session window and an all-models weekly limit; both only start counting after your first message.
• Trick: send a trivial "hi" to start the timer ASAP so it resets sooner, without burning usage.
• Automate keep-alive with a cron running `claude -p` every 5 hours (counts toward limits today, may change).
• Dual-wields two $200 accounts, swapping mid-session via `/login` + browser profiles (Helium). Account swap takes effect on the next tool call — lets you keep workflows running across accounts.
• Rough math: ~25% of weekly used per maxed 5-hour window → you can max the 5-hour window ~4x/week before exhausting weekly.

What to actually do with the tokens

• Run Claude Code workflows + ultra/ultracode mode for bulk parallel work (he spun up 100+ sub-agents auditing 3 competing PRs, hitting 1.6M+ tokens in <30 min).
• Daily ranked PR triage across repos (surfaces easiest/highest-value PRs as HTML plans; a month-old bugfix got merged in 5 min once surfaced).
• HTML plans over markdown — readable in browser/phone, pass URLs between Codex and Claude for cross-agent review.
• Run other people's audit skills (e.g. shadcn's improve skill) — skills are just markdown, copy raw and paste.
• Orchestration tip: Fable defaults to spawning Fable sub-agents; explicitly tell it to use Opus/Sonnet/Codex sub-agents if not deliberately burning money.
• Remote setup: Mac Mini with Codex + Claude Code + Hermes agent, accessed via SSH/Tailscale/screen share/T3 Code remote, so closing the laptop lid doesn't stop work.

The mindset

Don't token-max out of job fear — that just burns money and sanity. Approach with excitement to build things previously unimaginable: custom clouds, npm packages, forking your own tools. Lower the bar for what's worth building, raise the bar for how far you push. Also a caution: the "one more prompt" loop is wrecking maintainers' sleep.