Model Architecture

# Model Architecture Insights

<!-- Extracted and analyzed by Sayan Chowdhury (https://github.com/saynchowdhury) -->

This document analyzes what the Claude Fable 5 system prompt reveals about the model's architecture, hierarchy, tooling, and infrastructure. While system prompts are behavioral instructions rather than technical specifications, they contain substantial clues, analyzed here by **[Sayan Chowdhury](https://github.com/saynchowdhury)**, about the underlying engineering decisions and system design.

---

## 1. Model Hierarchy: The Complete Claude 5 Lineup

The system prompt reveals the full model hierarchy as of mid-2026. Listed from least to most capable:

- **Claude Haiku 4.5** (model string: `claude-haiku-4-5-20251001`) -- the lightweight, fast model
- **Claude Sonnet 4.6** (model string: `claude-sonnet-4-6`) -- the balanced mid-tier model
- **Claude Opus 4.8** (model string: `claude-opus-4-8`) -- the previously top-tier model
- **Claude Fable 5 / Claude Mythos 5** (model string: `claude-fable-5`) -- the new Mythos-class frontier

Several observations emerge from this hierarchy. The version numbers suggest different release cadences for different tiers: Haiku is at 4.5, Sonnet at 4.6, and Opus at 4.8, implying that Opus has received more point releases than the lighter models. The jump from the 4.x family to the 5 family for Fable/Mythos suggests a generational leap rather than an incremental update.

The Haiku 4.5 model string includes a date suffix (`20251001`, suggesting October 1, 2025), while the others do not. This is consistent with Anthropic's existing practice of including date-based version identifiers for some models but using simpler strings for newer releases. The prompt notes that "the person is able to switch models mid-conversation, so previous messages claiming to be from a different model or to have a different knowledge cutoff may be accurate," implying that the conversation platform supports dynamic model switching within a single session.

---

## 2. The Mythos-Class Designation

The "Mythos-class" designation is a new architectural category in Anthropic's model taxonomy. The prompt explicitly positions it: "part of a new Mythos-class model tier that sits above Claude Opus in capability." This is not simply a model name but a classification of capability level.

What makes a model Mythos-class? While the prompt does not define the criteria explicitly, several inferences are possible. First, the model has "dual-use capabilities" that require "additional safety measures" for general release. This suggests capabilities that could be misused at scale, possibly including advanced code generation, autonomous agent behavior, sophisticated reasoning about security systems, or other domains where high capability creates genuine risk.

Second, the model is described as the most capable in Anthropic's lineup, above even Opus 4.8. Given that Opus has historically been Anthropic's most capable tier (designed for complex reasoning, long-form content, and difficult tasks), a tier above Opus represents a significant capability threshold.

The Mythos-class designation may also carry architectural implications. It is possible that Mythos-class models use a different architecture, larger parameter count, longer context windows, or different training methodologies compared to the standard tiers. The separate classification suggests these models are not simply larger versions of existing models but represent a qualitatively different approach.

---

## 3. Fable vs Mythos: Same Model, Different Safety Configurations

The prompt states clearly: "Claude Fable 5 and Claude Mythos 5 share the same underlying model." This is architecturally significant because it confirms that the difference between these two products is not model weights or architecture but safety configuration.

Claude Fable 5 "includes additional safety measures for dual-use capabilities." This suggests a safety layer or filter that is applied at inference time, possibly through additional system prompt instructions, output classifiers, refusal fine-tuning, or runtime guardrails that intercept certain types of requests or outputs.

Claude Mythos 5 is "available without those measures to only approved organizations." The access control is at the organizational level: only vetted institutions can access the unrestricted version. This implies an enterprise or government sales channel with contractual safeguards.

The architectural implication is that Anthropic has developed a modular safety system where the same base model can be deployed with different safety profiles. This is more efficient than training separate models and ensures that safety updates apply uniformly. It also means that the "true" capability level of the model is what Mythos 5 offers; Fable 5 is a deliberately constrained version.

This approach is analogous to how operating systems ship with different security profiles for consumer and enterprise use, or how cloud services offer different feature sets based on subscription tier, except here the gating is about safety rather than features.

---

## 4. Model Strings and Versioning Scheme

The versioning scheme visible in the model strings reveals Anthropic's approach to model identity and lifecycle management: