The AI industry is entering a new phase — one where the most powerful models are no longer simply unleashed on the public, but carefully stratified by risk level before they leave the lab.
That is precisely what happened when Anthropic — the AI safety company founded by former members of OpenAI — released Claude Fable 5, its first Mythos-class AI model available to general users. At the same time, it quietly launched Claude Mythos 5, a more capable sibling that remains restricted to a small circle of vetted organisations through a programme called Project Glasswing. To understand why this matters, it helps to start from the beginning.
The Concept Behind It
What Is a Mythos-Class AI?
Anthropic uses internal tiers to categorise its AI models by capability. Mythos-class is the company’s highest tier — models designed for advanced reasoning, coding, and technical analysis that go far beyond answering simple questions. Think of it like the difference between a pocket calculator and a research-grade supercomputer: both do maths, but one can model nuclear reactions.
Until now, Mythos-class systems stayed inside restricted environments because their capabilities raised genuine safety concerns. Claude Fable 5 is the first time Anthropic has opened that tier to the general public — with important guardrails attached.
What Can Claude Fable 5 Actually Do?
Most AI models you may have used — including earlier versions of Claude — are built around the single-prompt model: you ask a question, the AI answers it. Fable 5 is built differently. It is designed for multi-step tasks, meaning it can plan a project, carry out its stages sequentially, and maintain context throughout — rather than “forgetting” what it was doing between steps.
In practical terms, this means a developer could ask Fable 5 to review a codebase, identify weak points, suggest fixes, and then write the corrected code — all as one sustained task, not four separate prompts. Anthropic says performance improvements are most visible as tasks increase in complexity, with the biggest gains appearing in technical and problem-solving evaluations. The company has highlighted coding and reasoning benchmarks where Fable 5 outperforms its earlier systems, though specific benchmark scores were not disclosed at the time of this writing.
How the Pieces Fit Together
The Two-Version Strategy
Rather than choosing between a full public launch and complete restriction, Anthropic took a middle path: release two closely related models with different access rules. Think of it like a pharmaceutical company releasing an over-the-counter painkiller and keeping a stronger prescription version behind the pharmacy counter. The underlying chemistry is similar; the controls around it are not.
Claude Fable 5 is available through Anthropic’s API (an Application Programming Interface — the technical doorway that lets software connect to an AI system) and enterprise subscription plans. Claude Mythos 5, meanwhile, is reserved for organisations already participating in Project Glasswing, Anthropic’s controlled-access programme for partners who can operate under monitored conditions.
Anthropic is explicit that both models share the same underlying technology. The difference is not in raw capability but in how each model responds to requests Anthropic classifies as sensitive or high-risk.
Why Cybersecurity Is the Critical Variable
The origins of the Mythos programme explain a great deal about why these restrictions exist. Earlier Mythos versions were developed with a strong focus on cybersecurity — specifically, identifying software vulnerabilities. Two months before the Fable 5 launch, Mythos Preview attracted significant attention in technology and security circles for exactly this ability.
The same capability that helps a security team find a flaw in its own systems can, in the wrong hands, show an attacker how to exploit it. Fable 5 handles this tension by applying layered safeguards. When a user’s request enters territory Anthropic considers sensitive, the model can restrict, limit, or block a response. In some cases, it may hand off the query to an alternative model with stricter safety controls, reducing the likelihood of misuse. Anthropic says this significantly limits Fable 5’s ability to complete offensive cybersecurity tasks compared with an unrestricted Mythos system — though it concedes some legitimate technical queries may receive reduced answers if the system flags them as sensitive.
There is a telling institutional logic in Anthropic’s dual-model release that the company has not made fully explicit: by publishing Fable 5 with documented safeguards while simultaneously restricting Mythos 5, Anthropic is effectively creating a public compliance record. At a moment when Anthropic’s CEO is calling for FAA-style federal oversight of powerful AI models, demonstrating a tiered, auditable access framework positions the company ahead of regulation rather than behind it — a strategy that could matter enormously if Washington moves toward mandatory AI capability disclosures.
How Claude Fable 5 Compares to Alternatives
Anthropic is not the only lab grappling with how to release powerful models responsibly. Here is how the major approaches compare on key dimensions:
| Lab / Model | Tiered Access? | Cybersecurity Restrictions? | General Public Access? |
|---|---|---|---|
| Anthropic — Fable 5 / Mythos 5 | Yes — two distinct release tiers | Yes — explicit safeguards on offensive tasks | Fable 5 yes; Mythos 5 vetted partners only |
| OpenAI — GPT-4o / o-series | Partial — API vs. consumer ChatGPT | Usage policies, but largely single model stack | Yes, via ChatGPT and API |
| Google DeepMind — Gemini Ultra | Partial — consumer vs. enterprise tiers | Usage policies applied at API layer | Yes, via Gemini apps and API |
| Meta — Llama (open weights) | No formal tiers — open download | Acceptable-use policy only; no technical enforcement | Yes — fully open weights |
Note: This comparison is based on publicly available information at the time of writing. Model capabilities and access policies change frequently.
The contrast with Meta’s open-weights approach is particularly sharp. As Anthropic has publicly warned about AI systems accelerating beyond human control, its decision to maintain hard technical restrictions — not just policy rules — reflects a fundamentally different philosophy from labs that treat openness as a safety feature in itself.
What People Get Wrong
Misconception 1: “Fable 5 is just a renamed, watered-down Mythos”
This is understandable but inaccurate. Anthropic says both models share the same core technology. The differences are in behavioural guardrails applied to specific request types, not in the model’s underlying reasoning engine. For the vast majority of professional and creative tasks, Fable 5 is genuinely the most capable Claude model ever released to the public.
Misconception 2: “Restricting Mythos 5 means it’s dangerous”
The restriction reflects caution about specific capability domains — primarily advanced cybersecurity analysis — not a general judgment that the model is unsafe. The same logic applies to why certain medical-grade diagnostic tools require a licensed professional: it is about context and accountability, not blanket danger. This nuance matters as the broader debate around what responsible deployment of Mythos-class models actually looks like continues to evolve.
Misconception 3: “These safety measures are foolproof”
Anthropic itself does not claim this. The company acknowledges that Fable 5’s safeguards reduce — not eliminate — the risk of misuse in sensitive areas. Layered technical controls are meaningful steps forward, but no AI safety mechanism to date has been proven impenetrable. Healthy scepticism is warranted, and the ongoing reality of AI-assisted cyberattacks is a reminder of the stakes involved.
Where to Learn More
- Anthropic’s Research page — primary source for model cards, safety evaluations, and technical documentation on Claude models.
- Anthropic’s Claude product page — for up-to-date access tiers, API details, and enterprise plan information.
- NIST AI Risk Management Framework — the US government’s framework for understanding and managing AI risk, useful context for why tiered access matters institutionally.
- For a broader look at where AI builder concerns are heading, why AI’s richest builders are starting to fear what they’ve built is a useful companion read.
The Implications That Matter
- Tiered AI access is becoming the industry norm, not the exception. Anthropic’s two-model approach signals a maturing recognition that not all capabilities should be equally accessible; other major labs will face growing pressure to adopt similar frameworks as models become more powerful.
- Cybersecurity is now the sharpest edge of the AI safety debate. The Mythos programme’s origins in vulnerability research mean that how Fable 5 handles offensive security queries will be scrutinized closely by both regulators and adversaries — the dual-use problem is no longer theoretical.
- Anthropic’s reported IPO filing changes the stakes for its safety commitments. A company moving toward public markets must balance investor appetite for growth with the credibility of its safety brand; Fable 5’s guardrails are simultaneously a product feature and a trust signal to future shareholders.
- Project Glasswing’s expansion will be the real test of this strategy. The vetted-partner tier is only as robust as its vetting process; as demand for Mythos 5 grows, the criteria and rigour of that programme will determine whether the tiered model holds or quietly erodes.
