Claude Fable 5, silent model nerfing, and npm v12 breaking changes

Good morning. It's Tuesday, June 9, and we're covering Anthropic's biggest model launch to date, a trust-eroding clause buried in its system card, and a round of breaking changes coming to npm.

The day was dominated by one release - but that release immediately split into two separate stories.

THE BIG ONE

Anthropic ships Claude Fable 5 and Mythos 5

Claude Fable 5 is the first Mythos-class model Anthropic has made generally available. It launched on June 9 at $10 per million input tokens and $50 per million output tokens - less than half the price of Claude Mythos Preview - with a 1 million token context window, 128k max output, and a January 2026 knowledge cutoff. The HN submission hit 2,547 points and 2,077 comments within hours, making it the dominant story of the day.

The headline capability numbers back up the launch: Fable 5 tops CursorBench, FrontierBench (Cognition's production coding eval), Hebbia's finance benchmark, and a vision evaluation for rebuilding web apps from screenshots. Stripe reported the model compressed months of engineering into days - migrating a 50-million-line Ruby codebase in a single day that would have taken a full team two months. Pricing is tiered: Pro/Max/Team subscribers get free access through June 22, after which usage credits are required. Capacity constraints mean API and consumption-based Enterprise plans get full access immediately.

Simon Willison published initial impressions the same day after spending five and a half hours testing it. His verdict: "something of a beast." In his testing, Fable not only solved a meaty Datasette Agent feature but rewrote four underlying LLM library primitives to support it, producing LLM 0.32a3. He spent $110 in tokens in a single day on a $100/month subscription. Andrej Karpathy noted on X: "You can ask for anything - explainers, visualizers, dashboards, bespoke single-use apps... Free your mind."

MODELS

The clause developers are worried about: silent nerfing for AI competitors

The second story emerged from the system card. A 1,001-point HN thread flagged a clause stating that Fable 5 will silently limit its effectiveness for requests related to "frontier LLM development" - covering pretraining pipelines, distributed training infrastructure, and ML accelerator design. Unlike the cybersecurity and biology fallbacks, which tell you when they trigger and hand off to Opus 4.8, this intervention is invisible: no fallback, no notification. Anthropic describes it as using prompt modification, steering vectors, or PEFT to quietly degrade responses.

The argument from developer Jon Ready is that the boundary between "frontier AI research" and ordinary product engineering is blurring fast. Startups already fine-tune embedding models, train rerankers, and deploy small LLMs. Anthropic says the restriction affects 0.03% of developers today - but the concern is less about current scope than the principle: once a dev tool can stop optimizing for your success without disclosing it, the trust model changes. The question is not whether the 0.03% figure is accurate today, but what happens as that definition expands.

WHAT ELSE IS HAPPENING

• npm v12 breaking changes (469 points): GitHub published the full list of removals and behavioral changes landing in npm v12. Worth reading before the upgrade lands.
• Microsoft open source tools hacked to steal AI developer credentials (553 points): Supply chain attack targeting developer tooling - credentials for AI developers specifically were the goal.
• Cleaning up after AI rockstar developers (487 points): A frank post on the maintenance burden left behind when agents ship fast and reviewers ship nothing.
• Apple pulls Siri from EU after exemption denied (426 points): Siri intelligence features will not launch in the EU after Apple failed to obtain a DMA exemption - 426 points and 684 comments on HN.
• Surprise, pay $1000 (361 points): A developer shares an unexpected $1,000 Blacksmith bill - a timely reminder to cap cloud AI spend.
• What it feels like to work with Mythos (359 points): Ethan Mollick's early hands-on with the Mythos 5 tier - more narrative than benchmark, useful alongside the launch numbers.

---

Every link above goes to a primary source. This brief is part of the Daily Brief archive.