Anthropic accidentally exposed a draft blog post and roughly 3,000 unpublished assets on March 27, 2026, via an unsecured public content cache. What was in there: a near-complete announcement for a model called Claude Mythos, described internally as “by far the most powerful AI model we’ve ever developed.” Anthropic confirmed development and testing after the leak without denying any of the specifics.
What Mythos Actually Is
Mythos sits in a new tier above Opus. The hierarchy is now Haiku, Sonnet, Opus, and then whatever comes next. Mythos is the first model in that new top tier. Capybara is almost certainly just the internal codename. There is no realistic world where Anthropic ships a product called Capybara given their naming scheme. Haiku, Sonnet, Opus, Mythos all follow the same logic around the weight and length of written form. Capybara does not fit. Mythos does.
Compared to Claude Opus 4.6, the leaked draft says Mythos gets dramatically higher scores in software coding, academic reasoning, and cybersecurity. No specific benchmark numbers leaked, but the draft language is not subtle. Dario Amodei has separately described it as a step change in capabilities. This is not an incremental Opus update. It is a genuinely larger and more capable model, and the new tier name signals that Anthropic is treating it that way.
For reference, you can see where Opus 4.6 sits relative to other frontier models in the full Claude launch timeline here.
The Cybersecurity Angle
The leaked draft spends a lot of time on cybersecurity, describing Mythos as far ahead of any other AI model in cyber capabilities. Anthropic’s release strategy is to give cyber defenders early access first, letting them use the model to harden codebases before the same capabilities are more broadly available. That framing is partly a real concern and partly marketing. A model this capable being used to find vulnerabilities is a genuine dual-use problem, but “our model is so powerful it’s scary” is also a reliable way to generate attention at launch. Both things can be true at once.
Cost and Who Will Actually Use It
Mythos is large and compute-intensive. Anthropic says it is very expensive for them to serve and will be very expensive for customers to use. There is no general release planned yet. The current rollout is invite-only, with early access going to a small group focused on cybersecurity applications.
For most use cases, Mythos is going to be overkill. Unless you are running an advanced cybersecurity operation or working on something extremely high-stakes like major legal analysis or critical infrastructure review, you are not going to need this model and probably will not be able to justify the cost even if you wanted to. The expensive frontier model matters less for day-to-day work than what it eventually teaches the cheaper models that follow it. When Mythos-level capability gets distilled down to Opus or Sonnet pricing, that is when the broad impact shows up.
Anthropic has been on a strong growth trajectory heading into this launch. The company hit $19B ARR recently, with major enterprise customers running production workloads on Claude. You can read more about that here. OpenAI also has a model internally called Spud that is supposed to be releasing soon, described as omnimodal with a meaningful capability bump. More on that here. Both labs are pushing hard on the next tier above their current flagships at the same time.
For now, Mythos is real, training is done as of March 2026, and it is coming. If you are not in cybersecurity or doing something that genuinely requires frontier-tier reasoning at scale, it is mostly a preview of where things are going rather than something you will touch anytime soon.