Anthropic Unveils Claude 'Fable 5,' Its First Public Mythos-Class AI Model: A Glimpse Of AGI, But Not There Yet

The competition to develop capable large language models (LLMs) has defined much of recent progress in AI.

OpenAI's release of ChatGPT brought conversational systems into everyday use and raised expectations for what models could achieve in reasoning, coding, and knowledge tasks. Other organizations responded with their own lines of models. Anthropic, founded by researchers who had worked at OpenAI and organized around goals of safety and reliability, built the Claude family through successive generations.

The Claude 3 models introduced a tiered structure with Opus, Sonnet, and Haiku variants.

Later releases in the Claude 4 series further improved performance in software engineering, agent-like workflows, and sustained reasoning over long tasks.

Anthropic's work on even more advanced systems, grouped under the Mythos class, produced capabilities that exceeded previous generally available models in several domains.

Early versions demonstrated particular strength in identifying software vulnerabilities across major operating systems and browsers. They also showed promise in supporting biological and chemical research workflows. These same strengths prompted internal caution at Anthropic. The company limited initial access to Mythos previews through Project Glasswing, a collaboration focused on defensive cybersecurity and critical infrastructure protection, often involving government partners.

The developers recognized that unrestricted release could create pathways for misuse in areas where the models offered meaningful uplift over existing tools.

Now, Anthropic announces Claude 'Fable 5,' its first Mythos-class model made available without those restrictions.

In the announcement, Anthropic said that:

Fable 5 uses the same core model weights as the more limited Mythos 5 but incorporates additional safeguards at the point of use.

Benchmark results place it at or near the top across most evaluated tasks.

Performance stands out in software engineering, where it handles extended autonomous sequences in large codebases, performs complex migrations, and meets production code standards with relatively high efficiency. Gains are also reported in knowledge work that involves multi-step analysis, document reasoning, and domain-specific judgment in fields such as finance and trading. Vision capabilities allow precise extraction of information from figures and diagrams as well as tasks like reconstructing interfaces from screenshots or operating environments with minimal additional scaffolding.

The advantages of Fable 5 grow more noticeable as tasks increase in duration and intricacy.

Testing showed it sustaining coherent work across millions of tokens while using its own generated notes or persistent memory to improve later outputs.

In scientific contexts it has produced novel hypotheses in molecular biology that internal experts often preferred over those from earlier models, and it has accelerated portions of protein design and genomics analysis pipelines. These results reflect improvements in planning, self-correction, and recovery from intermediate errors during long-running processes.

Anthropic developed the model through continued application of its Constitutional AI methods.

A written set of principles guides training and evaluation so that the system favors helpful, honest, and low-harm behavior without requiring exhaustive human review of every output.

For the public release of Fable 5, the company added dedicated safety classifiers.

These systems identify requests that touch on cybersecurity, biology, chemistry, or attempts to distill the model’s capabilities for use elsewhere.

When a classifier triggers, the request is routed to Claude Opus 4.8 instead of receiving the full Mythos-class response.

The mechanism activates in a small share of sessions on average and users receive clear notification. The classifiers were tuned conservatively to favor safety during the initial rollout, which can occasionally affect benign queries. Anthropic has stated it will continue refining them to lower false positive rates.

A parallel configuration, Claude Mythos 5, removes some of these restrictions and is being made available to a growing set of trusted partners through an expanded access program.

Initial focus remains on defensive cybersecurity work and biomedical research. Both versions are priced below earlier Mythos previews while delivering higher capability in permitted areas.

Internal teams at Anthropic that build coding tools have already adjusted their workflows around Fable 5.

Engineers report spending less time confirming that individual outputs are technically correct. Attention has shifted toward verifying that the model is pursuing the right overall objectives and breaking problems down appropriately.

This change arises because the model more consistently produces reliable intermediate results and sustains effort across multi-stage projects.

The release of Fable 5 shows one practical way to broaden access to frontier capabilities while addressing dual-use risks through layered controls.

Read: Even Anthropic Is Worried: AI May Be Advancing Faster Than Humans Can Manage, Prompting Calls for an AI "Brake Pedal"

It allows wider use in software development, analytical work, and research support without granting unrestricted access to the model’s strongest performance in the most sensitive domains.

At the same time, the approach underscores the technical effort required to evaluate, classify, and route requests at scale.

Continued progress in these systems forms part of the broader movement toward artificial intelligence that can reason and act across a widening range of complex, open-ended problems with greater independence.

How organizations and society manage deployment, oversight, and access for such models will influence both the pace of useful applications and the mitigation of associated risks.

But even with those advancements, and that people have started referring Fable 5 as showing a glimpse of AGI, things are actually far from it.

Fable 5 represents a significant advance in frontier large language model capabilities, with strong results in long-horizon software engineering, vision tasks, scientific hypothesis generation, and sustained multi-step workflows. These improvements allow it to handle more complex and extended projects than previous models with greater efficiency and autonomy in permitted domains. However, it remains a highly capable narrow AI system rather than a generally intelligent one.

However, "true AGI" would require reliable performance across a broad range of intellectual tasks at or above human expert level, including robust generalization to novel situations, consistent long-term planning without heavy scaffolding, and the ability to produce genuine novelty outside patterns seen in training data.

Current models, including Fable 5 and its contemporaries, are still fundamentally advanced Artificial Narrow Intelligence (advanced ANI, or "proto-AGI" in optimistic framings) that excel as sophisticated pattern matchers and synthesizers within their training distributions. They're not AGI because they still exhibit limitations such as hallucinations on unfamiliar problems, dependence on human oversight for high-stakes work, and the need for external tools or prompts to maintain coherence over complex sequences.

The release of Fable 5 with its safety classifiers and automatic fallbacks to weaker models like Opus 4.8 illustrates these boundaries in practice.

Such mechanisms exist because the underlying capabilities.

While powerful in areas like coding and research support, Fable 5 as Anthropic's current most powerful model, is not yet trustworthy or general enough for unrestricted use across all domains.

Regardless, Fable 5 and similar systems mark continued movement along the path toward more AGI rather than its realization.

The field is in an active phase of capability scaling combined with iterative work on evaluation, alignment, and deployment controls. The distance that remains on dimensions such as autonomous novelty generation and robust cross-domain reasoning keeps AGI as an ongoing objective rather than a current achievement.

Read: Paving The Roads To Artificial Intelligence: It's Either Us, Or Them