It doesn’t matter what technological product we are talking about, because both the product and how it is sold to you matters. And here making promises and generating expectations is the classic strategy. The next processor is going to be more powerful, the next smartphone is going to take better photos… and of course, the next AI model is going to be (much) better. We are seeing that message constantly in the AI segment, but now it is going further.
Anthropic and a curious leak. A group of security researchers they detected a few days ago 3,000 unpublished documents in an accessible Anthropic database. They included a draft of the blog entry that corresponded to the theoretical launch of their next AI model. The striking thing is not so much the filtration itself (whether intentional or not), but what those documents reveal.
Mythos goes beyond mere evolution. Or at least that’s what that leaked draft seems to reveal. It describes a model called Claude Mythos—also called Capybara—which would not be a simple improvement on Claude Opus, but would be a level above it. The document says that this model is “bigger and smarter than our Opus models, which until now were the most powerful.”
Anthropic signs up for hype. According to this leak, the benchmark scores would be notably higher than those of Opus 4.6 in programming, reasoning and cybersecurity. At Anthropic have ended up confirming the existence of this development, and have described it as “a level change” and “the most capable model we have created to date.” It’s not too surprising a phrase, because it’s basically the same thing they’ve been saying about every new model they’ve released.
And even they are scared. In fact, what is surprising in that draft is not the message that it is better, but the warnings that accompany that future presentation. Thus, Anthropic describes Mythos as “currently far ahead of any other AI model in cybersecurity capabilities.” In fact, they warn that this may be the beginning of “an imminent wave of models that can exploit vulnerabilities in ways that far exceed the efforts of the defenders.” Or what is the same: Mythos could be a extraordinary tool for cyber attackers. The actual launch plan is to first offer Mythos to cybersecurity organizations to prepare. We will see if that gives an advantage, if Mythos meets expectations.
OpenAI also makes a move. Both Anthropic and OpenAI have been moving in parallel for some time, and now they have done so again. At OpenAI they are preparing their new AI model, codenamed “Spud” (“potato”). Hardly anything is known about him beyond the fact that his pre-training phase has been completed. More relevant is that this model appears just when At OpenAI they have decided to be less OpenAI and more Anthropic. They have abandoned Sora and they are redirecting resources to regain ground where they are losing it. That is, in companies.
But the count is not infinite.. These days, users of Claude’s $100 and $200 per month plans began to notice how they used up their limits and token quotas in less than an hour during their work hours. What is happening is that Anthropic is training more powerful but much more expensive models to use and that makes it difficult to serve them. Demand is growing faster than the efficiency improvements that are coming, so according to some analysts, AI companies are adjusting those quotas and in a sense making Their models behave as if they were “dumber” to save. It’s something we’ve seen in the past.
hedonic adaptation. The psychologists called hedonic adaptation to the phenomenon by which humans quickly become accustomed to any level of experience, good or bad, and return to our starting emotional state. When applied to AI, this phenomenon explains that this model that seemed miraculous to us six months ago today seems slow and limited, and what six months ago seemed like science fiction is today the minimum we ask of companies. Anthropic and OpenAI have not invented the concept, but they have integrated it into their roadmaps like other technology companies in the past. We mentioned it before: they not only sell what they have today, but (more importantly) what they will have tomorrow.
Mythos will be brutal and very expensive. Anthropic’s draft warns that Mythos will be “very expensive to serve and will be very expensive for our customers.” That points to two possibilities. The first is that only users of the Max plans can access some consultations with this model. The second, that a subscription appears even more expensive than that 200 dollars a month so we can leverage Mythos with more leeway. We already had a free AI, a basic paid AI and a high-end paid AI. Now we will also have super high-end AI.
GIPHY App Key not set. Please check settings