Claude Mythos Preview: Benchmarks, Price, Context

Claude Mythos Preview is Anthropic’s unreleased multimodal model tier above Opus.

Anthropic has a new model listed as Claude Mythos Preview, and the listing is already doing the usual internet thing: turning a name, a few FAQ entries, and some benchmark chatter into a full-blown rumor mill. The model is described as proprietary, multimodal, and positioned above Claude Opus, with the internal codename “Capybara” attached to that higher tier.

That sounds vague at first glance, but the page gives enough signals to make a useful read. It points to Anthropic as the creator, says the model can take text and images, and frames it as an unreleased frontier system rather than a public product you can try today.

What the listing actually tells us

The LLM Stats page is light on hard product details, but it still gives a few concrete anchors. Claude Mythos Preview is listed as a multimodal model, which means Anthropic expects it to handle both text and image input. It is also marked proprietary, so this is not an open release with weights or a permissive license.

The more interesting part is the positioning. LLM Stats says Mythos sits above Opus, which matters because Opus is Anthropic’s top public Claude tier today. If that hierarchy is accurate, Mythos is less a small refresh and more a higher internal class of model that may eventually replace or sit beside the current flagship line.

• Created by Anthropic
• Proprietary license
• Multimodal input support
• Higher tier than Opus

Why the benchmark talk matters

The source text claims Claude Mythos Preview identified thousands of zero-day vulnerabilities across major operating systems and web browsers. That is the kind of claim that instantly raises eyebrows, because it suggests a model with unusually strong reasoning and offensive security capability, or a benchmark story that needs more context before anyone treats it as proof.

There is a big difference between a model being good at finding bugs in controlled tests and a model being broadly reliable in real-world security work. If Anthropic is testing this system against vulnerability discovery tasks, that would fit a pattern we have seen across frontier models: benchmark performance can spike on narrow tasks long before product teams are ready to expose the model widely.

“We’re not building models to be clever in a demo; we’re building them to be useful in the real world.” — Dario Amodei, Anthropic, in a 2024 interview with WIRED

That quote matters here because it captures Anthropic’s public posture. The company has repeatedly framed its work around utility, safety, and controlled deployment, which makes sense for a model that appears to sit above Opus and is still labeled preview rather than released.

How it compares with Anthropic’s current lineup

On the public side, Anthropic’s Claude family currently centers on models like Claude Opus, Claude Sonnet, and Claude Haiku. Mythos Preview does not have a public API page, pricing card, or docs page yet, which makes it different from the models developers can actually buy and ship with today.

That gap matters more than the name itself. A lot of model launches get overread because people focus on raw capability and ignore deployment details. For teams building products, the real questions are still the boring ones: when will it be available, what will it cost, what will the context window look like, and how stable will the API be under load?

• Opus, Sonnet, and Haiku are public Claude tiers
• Mythos Preview has no public pricing page yet
• No official context window is listed on the source page
• No public API docs are linked from the listing

What developers should watch next

For now, Claude Mythos Preview is more signal than product. The listing tells us Anthropic is testing a model class above Opus, and the security-related benchmark chatter suggests the company may be exploring stronger reasoning in domains where mistakes are expensive.

If Anthropic turns this into a public release, the first things to watch are pricing, rate limits, and context length. Those three numbers will tell developers far more than the codename ever will. A higher-tier Claude model only matters to engineering teams if it can justify its cost in code review, research, agent workflows, or security analysis.

My guess: the next real update will come from Anthropic itself, not from leaderboard pages. Until then, Claude Mythos Preview is best read as a signal that the company is already testing what comes after Opus, and the market will care less about the codename than about whether the model is worth integrating.

If you want a closer read on how model tiers are being positioned across vendors, see our related coverage on Claude Opus 4.5 and GPT-5.1.