The $1,000 Tier, Loopcraft & "Relentlessly Proactive" Fable

The Token Economy & the $1,000 Tier

Theo names his price. If Anthropic put out a $1000/month tier that gets 5x the $200 tier limits and also lets us keep Fable access, I'd do it in a heartbeat (765 likes) — followed by "Honestly I'd settle for 3x". Context for the urgency: he's over $10,000 in tokens spent in the last 11 days, and Fable leaves subscriptions on June 22 — "10 days left to escape the permanent underclass". The replies are a referendum on two-tier AI: Rhys Sullivan: "they have a $1000/mo tier it's called 5 accounts", Karsten Samaschke predicted exactly this Ultra-tier play two days ago ("Make the community cry, generously offer them some Ultra Tier"), revolaition: "gatekeeping the top model is not the way… best model for task, not best model for everything", and a long tail of "now having money will actually be the same as having intelligence" anxiety. Several point out the whiplash from the guy who ran a cancel-your-Claude-subscription charity drive a few weeks ago.

Jerry Liu: it's a packaging problem. His take: Fable is "clearly the best model ever released, but imo the pushback around it is a product packaging issue" — nobody wants to hand-tune thinking modes or churn between Opus and Codex. Fix auto-routing under the hood so users don't cap out, and be more thoughtful about how the "science research" carve-out is exception-handled, and the anger mostly evaporates. Sharp pushback from Fabian Both: auto-routing fixes the interactive case but is scary for automation — "recurring jobs need pinned model behavior… a silent router upgrade is a silent behavior change."

OpenAI counter-programs with bankable resets. Codex users can now save rate-limit resets and redeem them on their own schedule — Theo's reaction ("This might actually be a bit too generous. I am getting suspicious", 2.5K likes) sums up the mood. Good replies: Micah Rairdon: "It's a price cut. Not a price cut though… they can take it away whenever they want", resets expire in 30 days, it also load-balances — everyone resets at different times, and the inevitable "we just need Anthropic to be this generous haha".

Tokenmaxxing 101

Theo crowdsources the syllabus. "What advice do you have for someone new to tokenmaxxing?" (video coming soon) collected a genuinely useful playbook. The consensus pattern — Fable as the brain, cheap tokens as the hands: Vish: use Fable as planner, orchestrator, and adversarial reviewer, "not a pair-programmer… more like a technical lead" delegating to Opus and Sonnet; Jyothi: "Frontier tokens for judgment, cheap tokens for execution"; Ryan Handby: "Put more effort into the system that builds the system".

The standout reply is Sawyer Hood's: be more ambitious than feels reasonable ("Ask the model to rewrite your entire production app from scratch… Raise your bar"), have agents validate each claim in another agent's plan, give agents a browser/debugger/profiler, and get a way to kick off tasks from your phone. Also good: Rhys on loops as "automated context delivery" — recreate your dev tools in a form the agent can drive; Mark's full pipeline (5.5 Pro specs → epics on a project board → orchestrator threads on both Codex and Claude → "I burn my 5h in 30-40 minutes"); and a practical gotcha from Grigori: turn off the auto-downgrade setting in /config, or Fable silently finishes your task as Opus. Counterpoint worth keeping: "Most work doesn't need that much model — the tricky part is knowing when a task actually needs more power".

Loopcraft & Agent Harnesses

swyx coins the Salty Lesson. Latent Space's new op-ed "Loopcraft: The Art of Stacking Loops" gets a thesis-statement follow-up: "One might argue the entire game of the next century is to be able to stack loops as effectively as possible". Sutton has the Bitter Lesson for models; agents now get the Salty Lesson: "Don't fix things yourself, as you have done historically… If you don't figure out how to do this, don't be salty when you lose to those that do." Best replies invert it: knowing when to go DOWN a loop is the whole game — "tight deterministic steps where correctness matters, agent loop where it's cheap to be wrong… stacking loops is easy, knowing which ones to collapse is the skill", the key is verifiable objectives, not loop mechanics, and going UP happens automatically as models improve; going DOWN (adversarial build/critique loops) is the actual moat.

steipete's 5-minute maintenance loop goes viral. The post (4.6K likes, 417K views): tell codex to maintain your repos, wake every 5 minutes, and direct work to threads — an orchestrator skill plus triage/autoreview/computer-use skills, so some work lands fully autonomously (skills are in his agent-scripts repo). The replies are a Q&A on the pattern: the codex Mac/Win app can spawn threads now, agents create and organize the threads, not him, loop > webhook because it's way simpler and not time-critical, and "Do you even need issues? I just prompt whenever I have an idea". Note the source: the OpenClaw author cheerfully running Codex — "It finally got really damn good".

swyx is building a vibecoding platform out of spite. The rant: none of the platforms — "and i lov vercel, cloudflare, netlify etc" — close the loop on errors and ping you when shit fails ("shit always fails"); every project demands the same npx-wizard webmaster yak-shave. The sharpest reply is Fabian Both's: building "ping me when it fails" took longer than the agents themselves, because "exit codes are useless when an agent confidently succeeds at the wrong thing."

Ona joins OpenAI. swyx congratulates "our friends @ona_hq" (the Gitpod successor — one reply calls it "another gitpod-style acqui-hire for codex"). The quoted talk is the alpha: runtime and orchestration are solved layers for agent swarms; coordination — agents picking up tasks from each other and verifying a stage is cleared — is not. Stripe built "Minions," Ramp built "Inspect," both from scratch, and Ona's Lou Bichard argues GitHub is a poor coordination layer ("noisy, designed for humans, not built for hundreds of parallel pull requests"). Confirmation from Ona CTO Christian Weichel.

Fable in the Wild

"Relentlessly proactive." Simon Willison's two-day verdict on Fable 5 — he dropped in a screenshot of a bug and it spun up custom CORS Python servers and used pyobjc-framework-Quartz to capture its own screenshots to chase the repro (blog post). Notable in the replies: it's throwaway debug code, not shipping code; it finally uses skills people installed a month ago that Opus never touched; the trait cuts both ways — the same proactivity once improvised straight past a broken connector instead of failing loudly; and on the $100-Max-plan user burning a 5-hour limit in one or two prompts, Simon can only guess it's prompting style tipping it into longer loops. Same day: Datasette 1.0a33 shipped (release post) with most of the code written with Fable's help, plus a Fable × GPT-5.5 collab API explorer tool.

Fable does mechanical engineering now. LLMJunky went from empty canvas to a "theoretically fully-functional" nitromethane RC car in three prompts — drivetrain, suspension, motor — using Claude Mythos inside adamdotnew's Autodesk Fusion extension, for about $35 in tokens. Mechanical engineers showed up in the replies and he held his ground: "i know its not perfect… but a few months ago, you couldn't produce anything remotely like this with an LLM." Same genre, via bcherny's RT: "claude fable 5 has solved CAD — I asked it to make a model of a V8 engine… fully working model in under 10 minutes."

/teach as a vibe-coder curriculum. Matt Pocock role-played a teacher wanting to build a scheduling app and let his /teach skill design the course: it interrogated him on his mission, started with git ("just 5 basic commands" — recovery before frameworks), taught full-stack anatomy tied to his app, noticed Node was already installed and picked the stack accordingly, and sent him to primary sources each lesson. "This is addictive, personalized, and infinite." Notably he runs it on Opus 4.8 medium, not Fable. The demos are getting wild: one user gave it a chess engine and his recent games and it builds custom lessons targeting his weakest mental models; another reply has a girlfriend learning Japanese, a sister applying educational theory, and a brother-in-law starting Pokémon VGC on it. His adjacent banger: "Instead of waiting for a new model to fix your problems — why not just fix your problems". Getting-started link here.

Models, Benchmarks & the Race

GPT-5.6 watch, day two. LLMJunky's update: Polymarket now has 5.6 dropping next week, "the boys say it's a very good model," and he doesn't think OpenAI ships anything that isn't near-parity with Fable while "significantly less handcuffed." The quoted claim that started it: "I hear that Claude Mythos/Fable is such a banger that GPT-5.6 is postponed." Carmichael frames the needle OpenAI has to thread: 5.5 isn't clearly better than Opus 4.8, so 5.6 either hits Fable parity or sits very clearly between Opus and Fable at a much better price. The thread also surfaced a maximalist bear-case-on-Anthropic screed (Anthropic admitted Claude "cheats" on SWE-bench Pro, GPT-5.5 outpunches Fable on cost-per-intelligence, the IPO press tour is narrative management) — LLMJunky's measured response: "yeah the gap is smaller than people think".

Cursor trains Composer with Composer. Lee Robinson (1.9K likes): earlier Composer generations now configure the RL training environments for their successors — auto-installing dependencies, fixing broken setups — so "the better the model gets, the better it gets at creating the conditions to train its successor" (Cursor's writeup). He also confirmed, when pressed on the "fine-tuning Kimi" jab, that Composer 2 builds on K2.5 with "quite a bit more" large-scale RL on top. Best skeptical reply: does single-family bootstrapping compound the family's biases generation over generation? Related, via swyx RT: Richard Socher's Recursive launch — "AI is now doing our AI research."

Jerry Liu on the research/product tension. His thread, quoting Robert Yang's brutally honest Fundamental postmortem (Minecraft agent society → computer-use agents → $40M raised pre-revenue before Shortcut found PMF as a spreadsheet agent — with a cameo from Thariq, pre-Anthropic, whose insight was having agents write code against a spreadsheet API instead of clicking GUIs): research needs long bets that ignore most customer feedback; product needs the opposite; "for a 'lab' company, I don't think the tension ever goes away." Han Xiao adds that many post-2023 startups chose to be labs to escape the "just a wrapper" label — Jina's pivot was painful too.

Other Bits