06-reference

stratechery fable 5 anthropic alignment ai tiers

2026-06-10·reference·source: Stratechery·by Ben Thompson

"Fable 5, Anthropic Alignment, AI Tiers" — @benthompson

Why this is in the vault

This is the sharpest outside read so far on the model RDCO's COO agent literally runs on. Thompson covers the Fable 5 public release (the productized Mythos), dissects Anthropic's new guardrail regime — including a silent, undisclosed degradation mechanism — and lays out an emerging multi-tier AI market that has a concrete, dated consequence for RDCO: Fable 5 leaves subscription plans on June 23 and moves to usage credits. Verified: no third-party sponsor; subscriber-funded with house podcast links only.

The core argument

Fable 5. Anthropic released Claude Fable 5, a Mythos-class model now available to enterprise customers and paid subscribers, two months after gating Mythos to limited users. Capability is a step up — benchmarks reportedly 10%+ over Opus 4.8 — but the guardrails are heavy-handed. High-risk topics (cyber, bio) trigger a visible fallback to Opus 4.8; Thompson tripped it on his first try with an innocuous question about a Penn Medicine GLP-1/breast-cancer study, suggesting brute-force category blocking. Far more serious: per the system card, requests touching frontier LLM development (pretraining pipelines, distributed training infra, accelerator design) are silently degraded via prompt modification, steering vectors, or PEFT, with no fallback and no disclosure to the user. Anthropic estimates ~0.03% of traffic affected, concentrated in <0.1% of organizations. Thompson's objection is the precedent: Anthropic is making the model deliberately less helpful in a category "and not telling the user."

Anthropic Alignment. Thompson notes the Anthropic Institute's recursive-self-improvement safety report landed the week before launch and is cited in the system card; he reads the sequencing as safety-as-marketing, a tactic he traces back to the GPT-2 "too dangerous" playbook. He doesn't deny self-improving AI risk (his recurring line: in the original fable, the wolf eventually comes), but argues the nerfing has obvious business logic: deny coding/self-improvement capability to competitors. He extends his own Agents-Over-Bubbles harness argument (model+harness integration is where agent differentiation lives; profits flow to integrated parts of the value chain) to conclude that frontier labs must integrate forward into software, which makes Anthropic's competitive set "all of software." If Anthropic silently undercuts explicit competitors today, future competitors (i.e., anyone building software on Claude) should price in the same risk. His most memorable framing: Anthropic has achieved "true alignment" — internally, between commercial self-interest and sincere safety belief — so every business-optimal decision is experienced by employees as altruism.

AI Tiers. Two policy changes ride along: (1) mandatory 30-day data retention for all Mythos-class traffic on first- and third-party surfaces, safety-purposes-only for now, with logging of human access; opting out of retention is no longer possible if you want the best models. (2) Fable 5 is included on Pro/Max/Team/seat-based Enterprise plans only through June 22; from June 23 it requires usage credits, with restoration to subscriptions promised "when sufficient capacity allows." Thompson maps an emerging five-tier market: free/on-device baseline (Siri AI class) → cheap paid plans → serious subscriptions → serious API usage → and the top tier, Anthropic itself, whose internal devs get unrestricted Mythos and compound the lead. Speculative endgame: if the lead grows, retained data quietly becomes training data and customers go along. The only check he sees is competition — specifically whether OpenAI (GPT 5.5 already rumored better than Opus 4.8 by some) can surpass Mythos/Fable and force discipline.

Mapping against Ray Data Co

Strong mapping — this issue is about RDCO's own substrate.

Related