What Does 'Mythos-Class' Mean? Anthropic's Model Tier Explained

If you read Anthropic’s June 9, 2026 announcement and stopped on the phrase “a Mythos-class model that we’ve made safe for general use,” you probably had one question: what does Mythos-class actually mean? Here’s the short version. “Mythos-class” is not a product you buy and it’s not a separate model with its own API name. It’s the capability tier of the frontier model that sits behind two shipped products: Claude Fable 5 (the public, safety-on version) and Claude Mythos 5 (the same underlying model with some safeguards lifted, available only to vetted partners). One engine, two safety postures, one tier name. If you want the consumer-facing side of this, the Fable 5 product overview covers the model you can actually call today.

That distinction matters more than it looks, so let’s pin down exactly what the term describes, where it sits relative to Opus, Sonnet, and Haiku, and what it changes for you when you build against the API.

TL;DR

Mythos-class is the capability tier of the frontier model Anthropic put behind two products launched June 9, 2026. Claude Fable 5 is the public, safety-on version you can call. Claude Mythos 5 is the same model with safeguards lifted in some areas, restricted to vetted partners. Same engine, different safety postures, one tier name.

What “Mythos-class” actually means

Start with the exact wording Anthropic used. In the launch post, the company described Claude Fable 5 as “a Mythos-class model that we’ve made safe for general use.” Read that carefully. Fable 5 is described as a model belonging to a class. The class is “Mythos-class.” The phrase “made safe for general use” tells you that the raw class itself is not, by default, shipped to everyone.

So Mythos-class is a tier label for the underlying frontier model. It’s the most capable model Anthropic has ever put into general availability. Anthropic positions it as state-of-the-art on nearly all benchmarks, and it’s built specifically for long-horizon autonomous work, the kind of task that runs across millions of tokens without a human babysitting every step. Think multi-hour agent runs, large-codebase refactors, and research workflows that hold a lot of context at once.

The key mental model is the underlying-model framing. There’s one frontier model at this capability level. Anthropic then wraps it in different safety configurations and ships those configurations as named products. “Mythos-class” names the engine’s tier. “Fable 5” and “Mythos 5” name the two trims you can receive, depending on who you are and what you’ve been cleared for.

This is why you should not treat “Mythos-class” as a model ID. You won’t type it into an API call. It’s the category, the way “frontier model” or “flagship tier” is a category. What you actually invoke is claude-fable-5.

One more anchor for the name. The earlier preview of this same tier shipped under the name “Claude Mythos Preview.” Fable 5 and Mythos 5 are the productized, cheaper successors to that preview. We’ll come back to pricing, because the drop is steep.

Mythos-class vs Anthropic’s named models (Opus, Sonnet, Haiku)

Anthropic’s familiar lineup uses three names: Opus (most capable of the standard line), Sonnet (balanced), and Haiku (fastest and cheapest). Those are the generally available workhorses. Opus 4.8, for example, runs at $5 per million input tokens and $25 per million output tokens with a 1M-token context window. If you want the full breakdown on that model, the Opus 4.8 explainer and the Opus 4.8 pricing guide go deep.

Mythos-class sits above that standard line as a capability and codename tier. It’s surfaced to the public as Fable 5. So when you compare “Mythos-class” to “Opus,” you’re comparing a tier to a specific named model. The cleaner comparison is Fable 5 (the public face of Mythos-class) versus Opus 4.8 (the top of the standard line).

Here’s where each piece sits:

The practical reading: Opus 4.8 is the top of the everyday lineup and costs less. Fable 5 is the public on-ramp to a higher capability tier and costs more per token, because you’re paying for frontier-grade long-horizon performance. Anthropic’s models overview is the canonical place to confirm current IDs and pricing before you wire anything up.

Note the price relationship to the preview. Fable 5 and Mythos 5 both run at $10 per million input tokens and $50 per million output tokens. That’s less than half of what Mythos Preview cost. So the tier got cheaper as it productized, even as the capability ceiling moved up.

The two faces of the Mythos-class model: Fable 5 and Mythos 5

This is the heart of why “class” is the right word. Anthropic ships the same Mythos-class model in two safety postures, and the two postures are different products.

Claude Fable 5: safeguards on. This is the version made safe for general use. It’s publicly available, you call it as claude-fable-5, and it carries active safeguards. Specifically, Fable 5 routes a narrow set of high-risk queries to a different model. When a request touches cyber-offense, bio-chem weapons, or model distillation (attempts to extract the model’s weights or behavior into a clone), Fable 5 hands that query off to Opus 4.8 instead of answering with the full Mythos-class engine. Anthropic says this affects under 5% of sessions. So for the overwhelming majority of normal development work, you’re getting the full Mythos-class model. The routing only kicks in at the edges. The Fable 5 safety and safeguards breakdown walks through how that routing is triggered and what it means in practice.

Claude Mythos 5: safeguards lifted. This is the same underlying model with some of those safeguards removed in specific areas. It is not public. Access runs through a program Anthropic calls Project Glasswing, and it’s restricted to a vetted set of users: cyberdefenders, infrastructure providers, and select biology researchers. The logic is that defenders sometimes need the model to reason about the same threats that the public version deflects, so blanket refusals would hurt the people trying to harden systems. Mythos 5 narrows that gap for cleared partners while keeping the general public on Fable 5.

So the same engine appears twice. One trim is locked down and open to everyone. The other trim is unlocked in places and open to almost no one. If you want a side-by-side, the Fable 5 vs Mythos 5 comparison lays out the differences row by row. The thing to remember is that the capability tier, Mythos-class, is constant across both. Only the safety configuration changes.

Why the tier exists

The tier exists because two things were true at the same time: the model crossed into frontier capability, and frontier capability raises the stakes on misuse.

On the capability side, this is the most capable model Anthropic has put into general availability. It’s tuned for long-horizon autonomous work across millions of tokens, which is exactly the regime where small advantages compound. A model that can hold a giant codebase in context and act on it for hours is worth a lot. That’s the upside the tier is built to deliver.

On the safety side, the same long-horizon, high-capability profile is what makes uncontrolled access risky in a few narrow domains. Anthropic’s answer was not to pick one safety setting for everyone. It was to split the tier into two postures. The default posture (Fable 5) routes the riskiest queries away from the frontier engine, so the public gets frontier-grade help on normal tasks without frontier-grade help on weapons design or weight extraction. The restricted posture (Mythos 5) lifts some of those limits, but only for vetted defenders and researchers who have a legitimate need to reason about the threats.

That defensive framing is the point of Project Glasswing. The people most able to defend networks, infrastructure, and biological safety are often the people who most need an unfiltered model to do their work. Keeping them on the same locked-down trim as the general public would blunt their effectiveness. So the tier exists, in part, to make a clean separation possible: one engine, gated two ways, matched to two different risk profiles. None of this requires reading intent into Anthropic’s roadmap. It’s the structure described in the announcement itself.

What it means for you as a developer

For nearly everyone reading this, “using a Mythos-class model” means one concrete thing: you call Claude Fable 5.

Fable 5 is the public face of the tier. The model ID is claude-fable-5. You hit it through the same API surface you already use for Claude, so swapping it into an existing integration is a model-string change, not a rewrite. You get the frontier-grade, long-horizon capability of the Mythos-class engine, with the safety routing handling the small slice of queries that touch cyber, bio-chem, or distillation. If your workload is ordinary application development, agent orchestration, document analysis, or code generation, that routing won’t get in your way, because it fires on under 5% of sessions and those are the high-risk edge cases.

Mythos 5 is not something you can call. There’s no public endpoint, no self-serve access, and no API key that unlocks it. It’s gated behind Project Glasswing and limited to cyberdefenders, infrastructure providers, and select biology researchers. If you’re not in one of those buckets and cleared through the program, treat Mythos 5 as out of reach and build on Fable 5.

A few practical notes before you ship:

• Budget for the tier, not the standard line. At $10 input and $50 output per million tokens, Fable 5 is the priciest generally available model in the lineup, more than Opus 4.8. For high-volume or latency-sensitive paths, route those to Opus, Sonnet, or Haiku and reserve Fable 5 for the long-horizon, high-stakes work where the capability earns its cost. The Fable 5 pricing guide has the math.
• Confirm the model ID at integration time. Model strings and prices move. Check Anthropic’s models overview and the Anthropic news feed before you hard-code anything.
• Test the API contract, not just the prose. A new model can change response shapes, token accounting, streaming behavior, and tool-call formatting in subtle ways. This is where an API client earns its keep. With Apidog, you can build a request collection that hits claude-fable-5, assert on the response schema, snapshot token usage, and catch a breaking change before it reaches production. Set up an environment variable for your key, save the Messages endpoint, and run the same suite every time Anthropic ships an update.

If you’re standing up that test harness, Download Apidog and create a Claude environment with your base URL and key, then save Fable 5 calls as reusable requests so regression checks take seconds instead of a manual curl every time. The combination of a frontier model and a tight API test loop is how you adopt a new tier without surprises. You can also keep the same Apidog workspace pointed at Opus 4.8 so you can A/B the two models on cost and output quality for a given task.

That’s the whole concept. Mythos-class is the capability tier of one frontier model, shipped as Fable 5 for the public and Mythos 5 for vetted partners. If you build, your next move is to point your API client at claude-fable-5, run a quick request against the Messages endpoint, and confirm the response shape and token costs match what you expect before you wire it into anything real.