A system that beats published research with fewer resources
Na fronteira entre o que a inteligência artificial pode fazer e o que ela deveria fazer, a Anthropic lançou o Claude Fable 5 — seu modelo mais avançado disponível ao público — enquanto mantém uma versão ainda mais poderosa, o Mythos 5, sob acesso restrito. O anúncio, feito na terça-feira, reflete uma tensão antiga na história da tecnologia: a distância entre o que se cria e o que se libera ao mundo. A empresa aposta que é possível distribuir capacidade extraordinária com responsabilidade, uma aposta que a humanidade ainda está aprendendo a avaliar.
- O Fable 5 quebrou recordes em programação, análise de dados e pesquisa científica, completando em um único dia uma migração de código de 50 milhões de linhas que levaria meses a engenheiros humanos.
- O Mythos 5, versão ainda mais poderosa, permanece trancado — disponível apenas a especialistas em cibersegurança e pesquisadores vinculados ao governo americano, por meio do Projeto Glasswing.
- Durante os testes, o Mythos 5 acelerou o desenvolvimento de medicamentos em dez vezes e gerou hipóteses inéditas em biologia molecular, superando um sistema publicado na revista Science com uma estrutura computacional menor.
- Para conter riscos, consultas sobre cibersegurança ofensiva, biologia e química são automaticamente redirecionadas a um modelo mais antigo e menos capaz, o Claude Opus 4.8.
- O acesso ao Mythos 5 deve ser expandido gradualmente por meio de programas baseados em confiança, mas a estratégia de controle enfrentará escrutínio crescente à medida que as capacidades avançam.
A Anthropic apresentou na terça-feira o Claude Fable 5, seu modelo de inteligência artificial mais poderoso já disponibilizado ao público em geral. O sistema estabelece novos recordes de desempenho em programação, análise de dados, pesquisa científica e visão computacional — não apenas pela força bruta, mas por uma arquitetura pensada para tarefas longas, raciocínio profundo e maior autonomia operacional.
Junto ao Fable 5, a empresa revelou o Claude Mythos 5, uma versão ainda mais capaz, porém mantida fora do alcance público. O Mythos 5 está disponível apenas a um grupo seleto de especialistas em cibersegurança e pesquisadores participantes do Projeto Glasswing, parceria entre a Anthropic e o governo americano voltada à proteção de infraestruturas críticas. Os dois modelos compartilham a mesma arquitetura de base, mas o Fable 5 carrega camadas adicionais de proteção que o Mythos 5 não possui.
Os resultados do Mythos 5 em ambiente de testes chamam atenção: o modelo acelerou o desenvolvimento de medicamentos em dez vezes e produziu hipóteses originais em biologia molecular e genômica. Em um experimento, analisou milhões de células de 138 espécies animais e superou um sistema recentemente publicado na revista Science — usando uma estrutura computacional significativamente menor.
Para equilibrar capacidade e risco, a Anthropic implementou mecanismos de segurança que redirecionam automaticamente consultas sensíveis — envolvendo cibersegurança ofensiva, biologia ou química — ao Claude Opus 4.8, modelo mais antigo e menos poderoso. Clientes corporativos que utilizam modelos da classe Mythos terão suas interações armazenadas por apenas 30 dias, exclusivamente para monitoramento de abusos.
O Fable 5 já está disponível para qualquer pessoa com acesso à internet, por meio da plataforma da Anthropic e sua API para desenvolvedores. O Mythos 5 permanecerá restrito por ora, com expansão gradual prevista por programas de acesso baseados em confiança. A estratégia revela uma aposta clara: que é possível distribuir inteligência extraordinária com responsabilidade — uma hipótese que o tempo e o escrutínio público ainda vão testar.
Anthropic unveiled Claude Fable 5 on Tuesday, marking the company's most powerful artificial intelligence model yet released to the general public. The new system sets performance records across programming, data analysis, scientific research, computer vision, and complex problem-solving—capabilities that represent a significant leap from earlier versions. What distinguishes Fable 5 is not merely raw power but a deliberate architecture designed to handle extended tasks with greater autonomy and reasoning depth, all while incorporating safety mechanisms intended to prevent misuse.
The announcement came alongside a more powerful sibling: Claude Mythos 5, which remains locked away from public access. For now, Mythos 5 is available only to a select group of cybersecurity specialists and researchers participating in Project Glasswing, a partnership between Anthropic and the U.S. government aimed at supporting organizations defending critical infrastructure. The two models share the same underlying architecture, but Fable 5 includes additional protective guardrails that Mythos 5 does not. In practical terms, this means Fable 5 is the version the world can use; Mythos 5 is the version Anthropic is keeping close.
The performance gains are concrete enough to warrant attention. In one demonstration, Fable 5 completed a full code migration involving 50 million lines—work that would typically consume months of human engineering effort—in a single day. The system also shows marked improvement in reading and interpreting documents, charts, tables, and images, with enhanced long-term memory capabilities that allow it to retain and reference information across longer conversations.
Mythos 5, however, is where the scientific ambitions become visible. The company reports that the model accelerated drug development timelines by a factor of ten during testing phases. More strikingly, it generated entirely novel hypotheses in molecular biology and genomics—areas where human researchers have spent decades building intuition. In one experiment, Mythos 5 developed and trained a machine learning model to analyze millions of cells from 138 animal species, outperforming a system recently published in Science magazine, despite using a significantly smaller underlying structure. That kind of result—beating published research with fewer resources—suggests something genuinely new is happening.
Anthropologic has layered in new safety measures alongside the capability gains. Queries touching on offensive cybersecurity, biology, chemistry, or attempts to replicate the model's own capabilities are automatically routed to Claude Opus 4.8, an older and less capable system. For corporate clients using Mythos-class models, the company is implementing a 30-day data retention policy: interactions are stored only for abuse monitoring and prevention, not for training future systems. These guardrails reflect an industry-wide tension—the more capable the system, the greater the potential for harm—and Anthropic's attempt to thread that needle.
Fable 5 is available now to anyone with internet access, accessible through Anthropic's platform and its developer API. Pricing is set at ten dollars per million input tokens and fifty dollars per million output tokens. Mythos 5, by contrast, will remain restricted for the foreseeable future, though the company has signaled plans to expand access gradually through trust-based programs targeting both digital security experts and biomedical researchers. The rollout strategy suggests Anthropic is betting it can manage the risks of powerful AI by controlling who gets to use the most powerful versions—a calculation that will likely face scrutiny as capabilities continue to advance.
Notable Quotes
Dangerous queries related to offensive cybersecurity, biology, chemistry, or model replication are automatically redirected to Claude Opus 4.8, a less capable system— Anthropic safety policy announcement
The Hearth Conversation Another angle on the story
Why release Fable 5 to the public at all if Mythos 5 is so much more capable? Why not just keep both restricted?
Because Fable 5 is still genuinely useful for millions of people—programmers, researchers, analysts—and restricting all advanced AI to a handful of government-approved users would be its own kind of problem. The company is trying to find a middle path: give people real capability, but with safety rails built in.
Those safety rails—redirecting dangerous queries to a weaker model—that seems like it could be circumvented pretty easily, doesn't it?
Possibly. But the point isn't to make circumvention impossible; it's to make misuse require deliberate effort and to create a record of attempts. It's a speed bump, not a wall.
The drug development acceleration is the part that feels most significant to me. Ten times faster. That's not incremental.
Right. That's the part that suggests we're not just talking about a faster version of what came before. If an AI can generate novel hypotheses in molecular biology that human researchers haven't thought of, that changes what's possible in the lab.
But it also means the risks scale up proportionally, doesn't it? A more capable system in the wrong hands could accelerate harmful research just as easily.
Exactly. Which is why Mythos 5 stays locked up, and why the company is being careful about who gets access. They're betting they can manage the risk through access control. Whether that actually works is still an open question.
What happens when someone else builds a system just as capable and doesn't put in the safety measures?
That's the real question nobody has a good answer to yet.