Fable 5 Lands: Step-Change Praise, Safeguard Backlash & Nested Subagents

Claude Fable 5 Launch

The announcement. Anthropic introduced Claude Fable 5: "a Mythos-class model that we've made safe for general use. Its capabilities exceed those of any model we've ever made generally available." Claude Mythos 5 itself also shipped — same capabilities, without the safety classifiers, but not broadly available. Anthropic reset 5-hour and weekly rate limits for all users to celebrate (Theo: "THEY DID A RESET OMG").

Karpathy's endorsement (1.6M views, 20K likes): same underlying model as Mythos with added safeguards, SOTA on everything by a margin, but the qualitative jump is the story — "a major-version-bump-deserving step change forward (imo of the same order as Claude 4.5 was in November), peaking especially for long problem-solving sessions on very difficult problems… it's never felt this tempting to stop looking at the code at all (but don't do this in prod!)". He also concedes the safeguards are "a little too trigger happy for launch." The replies are a battlefield — more below.

Boris Cherny's take (535K views): biggest step up since Opus 4.5 — "Claude has stepped up from being a coding agent to a thought and design partner… Fable has judgement, taste, and dimensionality." His telling anecdote: it's the first model that debugs methodically — taking measurements, adding logs, verifying the fix before declaring victory — "There's nothing in claude code's prompting telling the model to do that, it's just part of its personality. It really has this 'big model smell.'" Anthropic's @trq212 agrees: "it's time to be more ambitious."

Simon Willison's initial impressions (blog post, via his post): "something of a beast — slow, expensive and capable of crunching through pretty much everything I threw at it." Key specs: 1M token context, 128K max output, January 2026 cutoff, priced at 2× Opus ($10/M input, $50/M output). The API has new mechanisms to signal when guardrails trigger, including an option to automatically fall back to another model on rejection. His knowledge-depth comparison vs Opus 4.8 is striking, and he posted pelican benchmarks per thinking-effort level with costs.

Other early reads:

Safeguards, Pricing & the Backlash

The counter-narrative is at least as loud as the praise, and it has several distinct threads:

Trigger-happy classifiers. Users report ordinary debugging and code review being flagged as "cybersecurity" and refused, or silently switched to Opus 4.8. A biotech founder is blocked entirely. Boris Cherny acknowledged it directly: "We know the classifiers are trigger-happy, and are working on improving it." kache's widely-liked take: "Silently nerfing itself is too much, it should at least give the user a message."

Intentionally degraded for frontier AI work. alphaXiv publicly objected to Anthropic "silently degrading Fable 5 for AI development" — pretraining pipelines, distributed training infra, and accelerator design "may have limited effectiveness." Karpathy's replies are full of researchers asking how they're supposed to use it (example), and Jeremy Howard's reply (2.7K likes) set the tone: "This is not a day for celebrating, Andrej. It's a very dark and very sad day, and the damage may be impossible to undo." When Theo asked him to elaborate, Howard answered: "Angry. And sad."

Data retention. Mike Bradley's viral summary of the objections: data captured and stored 30 days with no opt-out even for enterprise, no Zero Data Retention policy (a sore point for Bedrock customers), and capability held back for tiered offerings. His conclusion: "The best time to start learning how to self host was last year. The second best time is today."

The June 22 rug pull. Fable is included in Claude subscriptions only until June 22, after which it reverts to API-based billing. Jerry Liu: "We've got 13 days to burn as much tokens as humanly possible on Claude Max plans 💀". Theo: "Does kind of suck that we're going to be cut off from using it subsidized in like 2 weeks." Boris Cherny pushed back: "Our plan is to continue to offer Fable 5 as part of subscriptions, but we're being extra conservative because it's hard to forecast demand." Meanwhile swyx spots the arbitrage: "insane amounts of alpha in telling claude code to 'review my code for issues' on Fable rn while it is not pay per use… be prepared to be in abject horror that you shipped anything to prod without a Fable Check™ first."

Agentic Coding & Agent Harnesses

Nested subagents in Claude Code. Boris Cherny landed nested subagent support (413K views) — agents kicking off agents to manage context, capped at depth 5. Good detail in the replies: subagents can be monitored, can use Chrome, and his own experiment — add fork:true to a skill's frontmatter so it runs in its own context window, then have the skill use agents to isolate each step's context; he's adding this to the built-in /code-review skill. Sharpest skeptical replies: a silent failure at layer 2 propagates through layers 3–5 before surfacing (observability concern), and each subagent should return "a tiny receipt: what it tried, what changed, what it's unsure about — otherwise the parent agent inherits a mystery casserole". Related: his thread on self-verification loops as the key ingredient that lets models run longer.

Queues, not loops. Matt Pocock (186K views): "Everyone's banging on about loops, when they should be thinking about queues" — your issue tracker is a queue of tasks; agents pick tasks off it and completed work enters a different queue: human review. Replies worth reading: the bottleneck doesn't go away, it moves to the human-review queue that drains at the speed of one person reading diffs; David K Piano: "I'm thinking about state machines"; someone running Linear as an agent queue for an AI-research factory. Pocock also documented a full-day experiment: /to-prd → /goal with autoCompactWindow: 180000 to stay in the "smart zone" — Opus 4.8 immediately made the "classic AI mistake" of horizontal layers, but compaction held up and the verdict was positive.

pi ships Fable support. Mitsuhiko announced pi 0.79.1 — Fable support, a less annoying trust prompt, and more (release notes). Also a fun milestone: pi got mentioned in a WWDC 2026 talk (Apple video), plus pi lore: the original "pi" was Mario Zechner self-hosting Chinese models for a week when he feared Anthropic would take his Claude away.

Sandcastle + Codex. Eddy Vinck's setup guide for running Matt Pocock's Sandcastle sandboxed-agent runner against a Codex subscription (found via Pocock's RT).

Benchmarks & Evals

FrontierCode meets its first real test-time scaler. swyx on the Mythos/Fable launch: on FC Diamond, both Opus 4.8 and GPT-5.5 fail to meaningfully scale with effort — Mythos/Fable post-training is the first to convert test-time compute into solving "dozens of human hour equivalents, hundreds of dollars per task" problems. His rerun of historical models shows Fable "breaks every curve fit" across difficulty classes, and the easiest third of FC tasks went from 41% → 74% pass rate in four months in late 2025 — his data-driven explanation of the "WTF happened in Dec 2025" vibe shift. Also noted: 34 days between Anthropic signing the NVIDIA deal and a Mythos-class GA. Best reply: "if throwing more compute at a problem doesn't help, the model isn't reasoning — it memorized a neighborhood".

Fable thinks document parsing is beneath it. Jerry Liu's ParseBench results: Fable crushes reasoning-heavy benchmarks but on document understanding it's roughly equivalent to Gemini 3 Flash at 10–15× the token cost — and when asked, the model said it dislikes tasks "where the request is fully specified and the answer is fully known." Day-0 numbers from LlamaIndex: content faithfulness 90.02% (vs 86.19% Gemini 3 Flash) and a 12-point lead in semantic formatting — strong, but far short of specialized OCR providers (full leaderboard). Best reply: "it beats Runescape. it finds your invoice boring. same model." Related: LiteParse, their Rust doc parser, runs so fast Fable doesn't believe it's real (GitHub), and his earlier thread arguing agent filesystems are the new RAG.

Other Bits

  • OpenClaw plugin install policy. Big security update (via steipete RT): enterprises can register an executable that gets called with a full profile and source code of every plugin/skill before install is allowed — a gatekeeper hook for the plugin supply chain.
  • Claude Fable 5 Build Day, SF, June 13. Anthropic event with a $150K Claude-credits prize pool; @trq212 is meanwhile at Code w/ Claude Tokyo.
  • OpenCV 5 released (via LLMJunky RT): biggest update in years — new DNN engine with 80%+ ONNX coverage, built-in LLM/VLM support, performance that often beats ONNX Runtime.
  • Token accounting for new models. Simon Willison's TIL on using agentsview.io to track Fable 5 spending before the model lands in its pricing database.
  • IPO watch, continued. Simon Willison notes both OpenAI and Anthropic now have confidential S-1s with the SEC — Anthropic filed June 1. A reply in the swyx thread ties the launch to the filing: ~$10.9B projected Q2 revenue and a first-ever operating profit of $559M.