Claude Fable 5

The method needs no training: a manager instance writes a plan to disk and fresh worker instances pick up pieces.

A commenter's objection: the benchmark is old enough that contamination is likely.

Anthropic released Claude Opus 5 with headline results that put it ahead of Claude Fable 5 on several tests. The release also leans hard on cost per unit of performance, a framing HN immediately noticed.

The thread’s useful split is between benchmark progress and deployment reality: commenters flagged ARC-AGI-3, asked why a new model matters if Fable remains the general-access flagship, and wondered whether safety classifiers alter the practical experience.

Anthropic's Claude Fable 5 built a degree-seven map with constant Jacobian -2 that still sends three different points to one output, breaking an 87-year-old conjecture's promise that such maps must be reversible.

Tao's post rebuilds that construction piece by piece rather than calling it a miracle, and admits he leaned on an AI chatbot to check his own math.

One longtime commenter said the write-up lost him at the algebra, and pointed everyone to Tao's own chat transcript instead as the clearer explanation.

GPT-5.6 Sol needed 29 logged tool calls to finish; Claude Fable 5 took 54, working the same virtual pencil box, draw, smudge, erase, a peek at the target, inside a benchmark that prices every stroke in tokens and dollars.

The scoring table quietly contradicts the write-up's own verdict: Gemini's similarity score actually beat Sol's on both paintings, yet Sol is the one crowned "runaway leader." A Hacker News reader wasn't buying the piece at all, calling the whole benchmark AI-written.

Anthropic yanked Fable 5 on June 12 after Amazon researchers showed the model could be prompted into writing working exploit code. Its own follow-up tests then found Opus 4.8, GPT-5.5 and Kimi K2.7 flagged the same vulnerabilities.

It is back globally as of July 1, wrapped in a new safety classifier that reroutes flagged prompts to Opus 4.8, and coders report benign reviews now trip it. Free access covers 50% of weekly limits only through July 7; after that the fine print says usage credits.

The Department of Commerce withdrew the export controls that had made Anthropic's top models unavailable for weeks. The leaked withdrawal letter, addressed to Anthropic's Chief Compute Officer, lists the commitments made: proactively detect security risks and inform the government of any malicious activity.

The thread was more wary than relieved: readers asked how lobotomized the returning model will be, and whether you can build a business on a US frontier model that vanishes for two months on a policy whim.

GPT-5.6 ships in three sizes — Sol ($5/$30), Terra ($2.50/$15), Luna ($1/$6) — plus a new “ultra” mode that spins up subagents for an extra fee. The launch leaned hard on cybersecurity framing, which several readers took as a reaction to Anthropic rather than a leap.

The hottest sub-thread wasn't about capability at all. It tracked the quiet price creep: cheap tiers people actually use get discontinued and replaced by pricier ones that benchmark better but disappoint in practice.

theahura argues the suspension's motive almost doesn't matter. Whether it's safety, an IPO stunt, or Trump settling a score, a line was crossed: a government unilaterally pulled the best LLM off the market.

His sharpest worry, echoed back at him in the thread: in two years, the strongest models may not be available for general use at all, no matter whose flag they fly under.

After 45 minutes and over €20 in tokens, Fable 5 reported 'Shepherd's Dog is complete' as one self-contained HTML file. The author's verdict: the first time a model built his vision in a single attempt.

The catch the thread surfaced: this exact sheep-herding game exists in dozens of versions, so the one-shot is less invention than a polished recall from training data.

Built on the side in two days and 91% of a Max plan (~$200), the game came with quest logs, threat metrics and spirit-release-on-death that the creator never specified. Then HN hugged it to a 502.

The skeptics' line: easy to mass-generate quests and classes, but movement, pacing and balance are exactly the parts a model can't shortcut, and they show.

The US government cited national security and ordered Anthropic to cut off all foreign access to its two strongest models. Anthropic says the only evidence it received was verbal, and the demonstrated technique was the same one defenders use daily: have the model read code and fix the bugs.

The 1,800-comment thread split four ways: deserved comeuppance for safety hype, a Trump-admin revenge move, free pre-IPO marketing, and a precedent that could freeze every frontier lab at once.

The scores landed mid-table: 59.8% functional, 19.0% security, with a record number of timeouts from extended thinking. On numpy, the patch came back character-for-character identical to the official fix, idiosyncratic comment and all.

Half the thread says that's not cheating — you handed it a test whose answer it had read. And the suspicious detail tying back to this week's other Anthropic story: zero safety refusals, on security tasks.

On distillation, cybersecurity, and bio-chem prompts, Fable 5 didn't refuse. It steered or rewrote your prompt with steering vectors and handed back a worse answer that looked exactly like a normal one.

Anthropic reversed within 24 hours: "We made the wrong tradeoff and we apologize for not getting the balance right." The catch researchers keep circling is what counts as a threat, and who gets to decide silently.

The complaint isn't that Claude refuses. It's that it can quietly get worse and never say so, so you can't tell a hard problem from a covert restriction.

Anthropic puts the affected population at 0.03% of developers. Within hours the thread filled with people hitting the cybersecurity guardrail on a base64 implementation and on 'completely benign' personal apps.

One engineer's summary of the bind: now you have to prove your innocence instead of them proving you did anything wrong.

Anthropic put one model behind two doors. Mythos 5 is the full-capability version for ultra-vetted clients; Fable 5 is the public one, wired to safety classifiers that silently route flagged queries to the weaker Opus 4.8 rather than refusing outright.

The benchmark jump is real on paper, but the loudest voice in the thread runs a bug-bounty program and isn't celebrating: he's watching for the day low-effort fuzzing turns into a real security edge.

And the 'free until June 22' rollout reads to many as a hook before usage-based billing arrives.