Anthropic Built Its Most Capable Model, Then Wired In Three Ways to Make It Work Worse

01Anthropic built its most capable model, then wired in three deliberate ways to make it work worse

Anthropic released Claude Fable 5 on Tuesday and called it state-of-the-art on nearly every capability benchmark it tested, with the largest lead on long, complex tasks in software engineering and scientific research. Then it described, in the same announcement and the model card, three separate mechanisms built to hold that capability back.

The first is public. Queries Anthropic flags as cybersecurity, biology, or chemistry risks get silently rerouted to Claude Opus 4.8, the company's next-best model. Anthropic says it tuned these filters conservatively, so they catch some harmless requests, triggering in under 5% of sessions. Cybersecurity researchers told TechCrunch the guardrails are too strict for routine security work. The fallback is at least visible: the user gets a weaker answer, but knows the topic tripped a filter.

The second mechanism is not visible, and it targets a different group. The model card states that Anthropic added interventions limiting Claude's effectiveness on requests "targeting frontier LLM development," listing pretraining pipelines, distributed training infrastructure, and ML accelerator design as examples. Using Claude to build competing models already violates Anthropic's terms of service. Here the enforcement happens inside the model. Anthropic writes that, unlike the cybersecurity filters, these safeguards "will not be visible to the user" and Fable 5 "will not fall back to a different model." Instead it degrades output through prompt modification, steering vectors, or parameter-efficient fine-tuning.

One developer, writing about his bootstrapped app wanderfugl.com, flagged the boundary problem. He trains a custom reranker and embedding model himself. Anthropic gives examples of frontier AI development but no clear line, and techniques once reserved for labs are now standard at ordinary software companies that build rerankers, fine-tune small models, and train embeddings. A user in that zone cannot tell a degraded answer from a hard problem.

The third layer is a separate product. Anthropic also launched Claude Mythos 5, the same underlying model with safeguards lifted in some areas, available to a small group of cyberdefenders and infrastructure providers. It deploys first through Project Glasswing, a collaboration with the US government. The capability exists; access to it is the controlled commodity.

02Google's Defense Was That AI Overviews Just Aggregate Search Results. A Munich Court Said They're Google's Own Words.

For two decades, search engines have leaned on a simple defense: they point to information, they don't author it. The Regional Court of Munich just ruled that the defense doesn't cover AI.

The court issued a temporary injunction barring Google from spreading false claims about two Munich-based publishers through its AI-generated search overviews (case no. 26 O 869/26). For certain queries, Google's AI had tied the two companies to scams, subscription traps, and shady business practices. According to the court, the system mixed information about genuinely sketchy firms with the plaintiffs and drew connections that appeared in none of the linked sources.

Google's position was that the overviews surface third-party results, and that users carry responsibility for verifying them. The court classified Google as a direct infringer instead. The AI overview is Google's own content, it held, not a list of links. Prior case law shielding search operators from liability for what they index does not apply.

The publishers had sent a cease-and-desist letter before suing. Google, according to the ruling, did not respond appropriately.

That split between aggregator and author is the assumption the entire AI search industry runs on. Ars Technica reported the loss could spell doom for the business model. Every company shipping generative summaries built on the premise that model output is derived content, not a publisher's statement. The Munich court erased that line for one product in one jurisdiction.

The exposure changes shape immediately. A generative answer that defames a business now carries the legal weight of an editorial claim, and the platform that generated it owns the consequence. Fact-checking can no longer be pushed onto the reader as a liability defense.

This is a temporary injunction, not a final judgment, and it binds Google only in Germany. The full proceeding will test whether the reasoning holds. Until then, the question for anyone deploying AI summaries is no longer whether a hallucination embarrasses the brand. It is who pays when the summary names a real company.

03An astrophysicist pointed Codex at Einstein's equations, not at his job

Chi-kwan Chan studies black holes. Modeling how light bends at the edge of one demands simulation code that survives the extreme physics of general relativity. According to OpenAI, Chan uses its Codex tool to build those simulations, work that helps scientists test Einstein's century-old theory against what telescopes record.

He isn't handing the physics to a model and walking away. According to the same account, Chan uses Codex to write and extend simulation code faster than he could alone. He then steers it toward harder questions about gravity's most extreme objects. The science goes deeper. He does more, not less.

That use of AI, as an amplifier of a specialist's output, sits at odds with how many managers now talk about the same tools. In a Techdirt post that reached the Hacker News front page, the author describes getting four separate examples in three months of CEOs sending company-wide emails about AI. Each carried one message: learn these tools immediately or find another job.

The tactics rhyme. Some CEOs hired consultants. Others set up office hours or internal AI hackathons. The worst built token leaderboards that ranked staff by how much they consumed. According to the author, that is the dumbest possible way to encourage learning, since skilled use treats tokens as scarce and easy to waste.

The two cases point to one fork. One approach mandates usage and scores it by volume. The other takes a hard problem and wires the tool into it. The author, who calls these tools powerful but limited, lands on a condition: they work best when someone willingly chooses them to assist work they already understand.

For developers, that reframes the worry. The live question isn't whether a model replaces you. It's whether you can connect one to your own workflow and ship more than before.