Claude Mythos: The New Capybara Tier

Why a New Tier Matters

Since its inception, the Claude model family has operated on a clean three-tier structure: Haiku for speed and affordability, Sonnet for balanced performance, and Opus as the most capable flagship. Every Claude release slotted into one of these three tiers. With the arrival of Claude Mythos, that structure has fundamentally changed for the first time.

According to leaked Anthropic draft materials, "Capybara is a new tier name: bigger and smarter than our Opus models — and Opus has been our most powerful model until now." This is not an incremental Opus upgrade, not an Opus 5. It is an entirely new level of capability that demanded its own designation. The Capybara tier sits above Opus and represents the most powerful, most expensive class of model Anthropic offers.

An important terminology distinction: Mythos is the product name — what customers and developers will interact with. Capybara is the tier name — the structural category it occupies within Anthropic's model lineup, just as "Opus" is both a tier name and part of individual model names like Claude Opus 4.6. The underlying model is the same; the names serve different purposes in Anthropic's naming hierarchy.

The Updated Four-Tier Structure

Below is the complete Claude model lineup as of March 2026. The addition of Capybara expands the tier count from three to four, giving developers and enterprises a broader spectrum of cost-performance trade-offs.

Pricing for Opus, Sonnet, and Haiku reflects published API rates as of March 2026. Capybara pricing has not been disclosed.

What Makes Capybara Different from Opus

It is tempting to view Capybara as simply "the next Opus." That framing fundamentally mischaracterizes what Anthropic has built. Claude Mythos is not Opus 5 and should not be understood as an incremental upgrade within an existing tier. The distinctions are structural.

Scale. Claude Mythos is described as much larger than Opus models. While Anthropic has not disclosed parameter counts, the leaked draft materials and Anthropic's own public statements emphasize that the model is substantially bigger — requiring significantly more compute to train and to serve. This is not a fine-tuned Opus or a distilled variant; it is a fundamentally larger architecture.

Intelligence. Leaked benchmarks and Anthropic's characterization describe "dramatically higher scores" across software coding, academic reasoning, and cybersecurity compared to Claude Opus 4.6. Where Opus represents the upper range of the previous generation, Mythos represents what Anthropic calls a "step change" — a qualitative jump rather than a percentage-point improvement.

Cost. Anthropic has been unusually candid about the economics: Claude Mythos is described as "very expensive" for both Anthropic to operate and for customers to use. The company has stated that it is actively optimizing the model's efficiency before wider release. Community estimates for Capybara-tier token pricing range from 2x to 5x or more above current Opus rates, though no official numbers have been published.

Unknown specifications. Unlike the current model lineup, key technical specifications for Mythos remain undisclosed. Context window size, maximum output token count, and supported input modalities have not been publicly confirmed. For reference, the current models offer the following:

Pricing Context and Cost Optimization

Given that Capybara-tier pricing is expected to be significantly higher than Opus, cost management will be critical for teams integrating Claude Mythos into production workflows. Fortunately, the current Claude platform already offers several mechanisms for reducing per-token costs, and these are likely to apply to Mythos once generally available.

Batch Processing (50% Discount)

The Claude API supports batch processing for workloads that do not require real-time responses. Batch requests are processed within a 24-hour window and receive a 50% discount on token pricing. For high-volume analytical tasks — such as large-scale code review, document summarization, or data extraction — batch mode can cut costs in half. If Capybara-tier pricing reaches the levels community estimates suggest, batching may become essential for cost-conscious deployments.

Prompt Caching

Prompt caching allows developers to cache frequently used system prompts, context documents, and other static content so that repeated API calls do not re-process the same tokens at full price. For applications with long system prompts or shared context across many requests — common in agent-based architectures and multi-turn workflows — caching can dramatically reduce the effective input token cost. This feature is especially valuable at the Capybara tier, where every input token carries a premium.

Fast Mode (Priority Processing)

On the other end of the spectrum, Anthropic offers a Fast mode for latency-sensitive workloads. For Claude Opus 4.6, Fast mode carries a 6x pricing multiplier in exchange for prioritized processing and lower latency. While it remains to be seen whether a similar option will be available for Mythos, organizations with time-critical inference needs should anticipate a premium tier within the premium tier.

A Note on Naming and Versioning

Community discussions frequently use terms like "Claude Mythos 5" or "Mythos 5," positioning the model as the next generational leap after the Claude 4.x series. This framing is understandable — Mythos clearly represents a step beyond Opus 4.6 and Sonnet 4.6 — but it is important to note that Anthropic has not officially assigned a version number to Claude Mythos. The "5" designation is community-originated shorthand, not an official label.

Similarly, the relationship between tier names and product names can be confusing. In the existing lineup, the tier name and the product name overlap: "Opus" is both the tier and the model brand. With Capybara, Anthropic has diverged from this pattern. The tier is Capybara; the product is Mythos. Whether future Capybara-tier models will also carry the Mythos name, or whether Capybara will become a broader family like Opus, remains to be seen.

Platform Availability

The current Claude model family is available across multiple deployment channels, giving enterprises flexibility in how they integrate AI capabilities into their infrastructure.

Mythos deployment channels beyond the Claude API have not been disclosed. Given the model's computational intensity and the phased rollout strategy, availability on third-party cloud platforms may lag behind the direct API offering.

What the Capybara Tier Means for the AI Landscape

The introduction of a fourth tier signals something broader than a single model launch. It suggests that Anthropic sees a permanent need for a capability class above Opus — not as a one-off flagship, but as an ongoing tier that will receive future models. This has several implications for the industry.

First, it raises the ceiling on what "frontier AI" means in practice. If Capybara-tier models deliver the step-change performance Anthropic describes, competitors will face pressure to match not just the capability but the tier structure — offering their own ultra-premium models for tasks that demand maximum intelligence regardless of cost.

Second, it creates a more granular cost-performance spectrum. Developers previously chose between three tiers; now they have four. The right model for a given task depends not just on required quality but on budget, latency, and volume constraints. Capybara will likely be reserved for the highest-stakes applications — cybersecurity analysis, complex multi-step reasoning, mission-critical code generation — while Opus, Sonnet, and Haiku continue to serve the vast majority of production workloads.

Third, the pricing dynamics of a "very expensive" tier may accelerate innovation in efficiency. Anthropic has already stated it is optimizing Mythos before broader release. Techniques like distillation, speculative decoding, and hardware-level optimizations will be essential to making Capybara economically viable at scale — and those innovations will likely trickle down to benefit the entire model lineup.