frustrating

When Anthropic presented Claude Mythos Preview two months ago, he did it with a singular message: it is so powerful that you will not be able to use it. That, of course, caused everyone to want access to it. Well: Anthropic has just introduce Claude Fable 5 and Claude Mythos 5its new AI models directly derived from that. There is good news, but also bad news. Like Mythos, but capped as a precaution. Anthropic already warned that Claude Mythos Preview was a spectacular tool for finding security vulnerabilities. That made it especially juicy for cybercriminals, so the company decided that only a few trusted entities (under its Project Glasswing) would have access to the model. That learning has now been applied, because in this announcement we have two different (and layered) versions of the model: Claude Fable 5: a model with all the capabilities of Mythos Preview, but with notable security measures that prevent it from being used for malicious purposes. As soon as the model detects that we are asking something “dangerous”, it avoids the question and even forces the use of an inferior model, Claude Opus 4.8. Clear examples: questions about cybersecurity or the development of biological weapons, for example. Claude Mythos 5: This version is somewhat less capable than Fable 5 in terms of cybersecurity, but will only be available to “a small group of cyber defenders and infrastructure providers.” It is the natural heir to Mythos Preview, and according to its creators it is even better than the original version. Claude Fable 5 / Mythos 5 simply sweeps the most demanding benchmarks on the planet. There have never been more powerful models. Anthropic’s internal testing shows that we are facing the most powerful AI models in history. In all benchmarks – including the new FrontierCode programming, much more demanding than SWE Bench Pro – the scores of Claude Fable 5 and Mythos 5 are simply spectacular, well above those of their rivals. The jump from Claude Opus 4.8 is really surprising, but it leaves GPT-5.5 and Gemini 3.1 Pro far behind (they don’t compare with the recent 3.5 Flash). This is a brutal blow to Anthropic’s table, and we will see how both OpenAI and Google respond. Claude Fable 5 is amazing. Ethan Mollick, well-known AI popularizer, has had access to Fable 5 for a few days and is amazed by the experience. With this model he has managed to complete projects such as east of the isochronic map that previous models had never solved, and it has done it almost “the first time”. In one of the cases Fable 5 worked for 9 and a half hours straight to produce a code called Concord of data analysis. Their conclusions are compelling: Last year (when working with GPT-5 Pro) I called him “work with a magician”: you recite the spell and something happens. With Fable, the spell has become so powerful that I’m no longer sure I’m the wizard. I feel more like a patron. I describe what I want, pay for it and evaluate the result. The conspiracy takes place somewhere I can’t see, in hundreds of small decisions over which I never have a say. Work has gone from being a process to being a result. I no longer direct; charge. The criticism is unanimous. Andrej Karpathy, who recently signed by Anthropic, commented on X how this is a qualitative leap that for him is of the same relevance as the one that Claude 4.5 represented in November. That model began the overtaking of OpenAI: this puts it even further away (at least, for now). Other tweetersemployees or not from Anthropic, make it clear that this is an important leap in the capabilities of AI models. It’s only been a few hours since the launch, but everything points because we are indeed facing a notable leap in quality. Consume tokens like there’s no tomorrow. But in the face of that fascination, the criticism. Discussions on Reddit reveal how users who have started using it have quickly detected the problems associated with this release. The first of them: Claude Fable 5 burns tokens like there is no tomorrow. Its consumption is enormous, and the quotas for Pro and even Max accounts run out in minutes if we use the model intensively. If it already seemed to us that we were exhausting the limits of the free or quick payment accounts, with Claude Fable 5 that feeling worsens: Fable 5 is fantastic, but we can barely use it often with the Pro or Max plans because those dreaded messages about waiting X hours to continue using it quickly appear. Extremely cautious. Anthropic has been very serious about avoiding misuse of Fable 5, and as soon as it detects anything suspicious it “brakes” and “downgrades” the model so that at that moment the one that is activated is Claude Opus 4.8 (which is not bad at all). The problem is that users are detecting that the model takes completely harmless prompts as dangerous. Although in Anthropic indicate Although these security measures are activated in less than 5% of sessions, what users are detecting is that they are activated much more. Fable 5 can get silly. Not only that: Fable 5’s own design means that if it encounters a prompt that it detects as dangerous, the model tries to avoid the response and automatically reduces your capabilities (‘nerfing’) without you knowing. It gets a little sillier on purpose, so to speak. As Anthropic itself explains on the system card, We have implemented new measures that limit Claude’s effectiveness in requests related to the development of cutting-edge large-scale language models (LLMs) (for example, in creating pre-training pipelines, distributed training infrastructure, or designing machine learning accelerators). Using Claude to develop competing models already violates our Terms of Service, but enforcing this restriction through our security measures prevents giving an advantage to those users most willing to violate those terms. Unlike our cybersecurity, biology and chemistry interventions, and distillation attempts, … Read more