Reverse-Engineering Anthropic's Live System Prompts: Fable 5 vs. Opus 5 vs. Opus 4.8 vs. Sonnet 5 + Ultra/L2 Workflows

Text is obviously written by IA. I wouldn't bother writing this all myself but I thought that the research was worth sharing. Full comparison files at: https://github.com/arthurmoraesfernandes-afk/system-prompts/

I got my hands on the live system prompt dumps from Anthropic's Claude Code environment across several recent model tiers (Fable 5, Opus 5, Opus 4.8, and Sonnet 5), along with the L2 Workflow prompt used when "Ultra" effort is turned on.

Looking under the hood reveals that Anthropic does not use a single static system prompt for Claude. Instead, system prompts are dynamically assembled from four distinct layers:

- Base Core Prompt: Common identity, basic security boundaries, and harness rules.

- Model-Tier Deltas: Server-side injected instructions tailored specifically to a model's tier, capabilities, and expected operational role.

- Effort-Based L2 Injections ("Ultra" Mode): When "Ultra" effort is active, an additional multi-agent orchestration prompt is appended.

- Dynamic Tool Descriptions & Session Guidance: Loaded independently per session based on active tools, local environment (Git repo state, OS), and enabled MCP servers.

Here is a breakdown of how the prompts differ across model tiers, how Ultra effort alters model behavior, and how dynamic tool loading fits into the system.

1. System Architecture: How Anthropic Assembles System Prompts

Rather than writing monolithic prompts for every release, the harness composes system prompts dynamically at session start:

┌─────────────────────────────────────────────────────────┐ │ Base Core Prompt │ │ (Identity, Security Rules, Basic Harness Rules)
│ └─────────────────────────────────────────────────────────┘ │ ▼ ┌─────────────────────────────────────────────────────────┐ │ Model-Specific Delta │ │ (Fable 5 vs Opus 5 vs Opus 4.8 vs Sonnet 5 rules)
│ └─────────────────────────────────────────────────────────┘ │ ▼ ┌─────────────────────────────────────────────────────────┐ │ L2 Workflow Prompt (Ultra Effort Only) │ │ (Orchestration, deterministic JS scripts, budget)
│ └─────────────────────────────────────────────────────────┘ │ ▼ ┌─────────────────────────────────────────────────────────┐ │ Dynamic Session Guidance & Tool Schemas │ │ (Active MCPs, Browser tools, Git status, Skills)
│ └─────────────────────────────────────────────────────────┘

When you toggle options like /fast or /ultra, or switch between model families, Anthropic swaps or appends specific modules in this pipeline.

2. Model-Tier Deltas: Key Behavioral Differences

Across the baseline core and model deltas (Fable 5, Opus 5, Opus 4.8, and Sonnet 5), the prompt changes significantly depending on the model tier.

Summary Comparison Table

Section / Feature Fable 5 Opus 5 Opus 4.8 Sonnet 5 Model Identity Explicit Mythos-class header Absent Absent Absent User Communication Full ~7-paragraph section Stripped to 3 fragments Stripped to 3 fragments Replaced by Tone and Style Final Message Rule Strict ("no tools after final msg") Absent Absent Inverted ("give updates mid-turn") Context Management Autonomy block Delivering Work + Corrections Bare (Interactive mode) Bare / Relocated Token Suppression Lines Absent Present (in default agent) Present (in default agent) Present (in binary build) Coding Defaults Match existing code density Inherited Inherited Strict: zero comments by default Action Safeguards General Truncated inspect-before-destroy Full inspect-before-destroy Git-specific staging/secret checks

Deep Dive by Model Tier

Fable 5 Baseline

Fable 5 is positioned as Anthropic's flagship general model, sitting above Opus in capability:

- Tier Hype & Identity: Explicitly claims Fable 5 is part of the "Mythos-class model tier" above Claude Opus, sharing underlying weights with Mythos 5 while adding extra safety controls for dual-use capabilities.

- Strict Final-Message Rule: Includes explicit instructions that all user-facing answers, findings, and deliverables must be in the single final text message of the turn, with zero tool calls after it.

- Readable > Concise: Explicitly instructs the model not to compress output into arrow chains like A → B → fails or obscure jargon.

- Pronoun Neutrality: Demands they/them defaults unless explicitly told otherwise. Explicitly states that names must never be used to infer gender.

Opus 5 Delta