Claude Fable 5 Developer Guide: Agentic Coding at Scale

When Anthropic released Claude Fable 5 on June 9, 2026, it triggered over 237,000 posts on X in under 48 hours — the kind of signal that tells you something has genuinely shifted. This isn't incremental. Fable 5 is the first widely released Mythos-class model: same weights, same capabilities, built for the 95% of developer workflows where Mythos Preview was unreachable. This guide is your map for putting it to work.

What Makes Fable 5 Different

Claude Fable 5 (claude-fable-5) is Anthropic's most capable widely released model. It runs adaptive thinking by default — no separate extended thinking toggle needed — and supports a 1M token context window with 128k max output. At $10/$50 per million input/output tokens, it costs less than half what Mythos Preview charged, while matching it on 95%+ of tasks.

The benchmark gap over its nearest rival is not marginal. On SWE-Bench Pro — the industry's hardest real-world software engineering evaluation — Fable 5 scores 80.3%, compared to 69.2% for Opus 4.8, 58.6% for GPT-5.5, and 54.2% for Gemini 3.1 Pro. On SWE-Bench Verified it hits 95%. On FrontierCode (Cognition's evaluation for frontier coding difficulty), it leads every competing model at medium effort.

The gap widens as tasks grow more complex — exactly the range where autonomous agents operate.

Getting Started: API Setup

Fable 5 is available on the Claude API, Amazon Bedrock, Vertex AI, and Microsoft Foundry from June 9, 2026.

Python (Claude API):

import anthropic
 
client = anthropic.Anthropic(api_key="your-api-key")
 
response = client.messages.create(
    model="claude-fable-5",
    max_tokens=8192,
    messages=[
        {
            "role": "user",
            "content": "Refactor this Python module to use async/await throughout."
        }
    ]
)
print(response.content[0].text)

TypeScript (Node.js):

import Anthropic from "@anthropic-ai/sdk";
 
const client = new Anthropic({ apiKey: process.env.ANTHROPIC_API_KEY });
 
const msg = await client.messages.create({
  model: "claude-fable-5",
  max_tokens: 8192,
  messages: [
    { role: "user", content: "Generate unit tests for the UserService class." }
  ],
});
console.log(msg.content[0].text);

Amazon Bedrock (Python):

import boto3
import json
 
bedrock = boto3.client("bedrock-runtime", region_name="us-east-1")
 
response = bedrock.invoke_model(
    modelId="anthropic.claude-fable-5",
    body=json.dumps({
        "anthropic_version": "bedrock-2023-05-31",
        "max_tokens": 8192,
        "messages": [{"role": "user", "content": "Review this PR for security issues."}]
    })
)
result = json.loads(response["body"].read())
print(result["content"][0]["text"])

The same claude-fable-5 model ID works on Vertex AI and Microsoft Foundry without platform-specific prefixes.

Agentic Features: Ultra Code and Dynamic Workflows

Fable 5's headline capability is not raw benchmark scores — it's its behavior on long-horizon autonomous tasks. Two features define its agentic profile:

Ultra Code: Parallel Sub-Agents

Ultra Code spins up multiple sub-agents that execute tasks concurrently. When working on a large codebase with independent modules, Fable 5 can assign each sub-agent a separate component — tests, documentation, refactoring, security review — running them simultaneously. The orchestrator model synthesizes results and resolves conflicts.

Stripe reported compressing two months of manual team effort into a single day on a 50-million-line codebase migration. That scale of gain becomes baseline on sufficiently complex projects.

Dynamic Workflows

For highly dependent, multi-stage projects — where stage N requires the exact output of stage N-1 — Dynamic Workflows forces Fable 5 to generate a linear plan with explicit evaluation checkpoints before executing. This prevents the common failure mode where an autonomous agent plows through a chain of assumptions, only to backtrack several steps when it hits a validation wall.

File-Based Memory for Multi-Day Sessions

Fable 5 achieves a 3x performance improvement on complex stateful tasks when given persistent file-based memory. In practice, this means writing intermediate findings to files the model reads back on subsequent turns. For multi-day sessions in Claude Code, this pattern is the difference between coherent progress and context drift.

Use the effort parameter to control thinking depth per request:

response = client.messages.create(
    model="claude-fable-5",
    max_tokens=16000,
    extra_body={"effort": "high"},  # or "medium", "low"
    messages=[{"role": "user", "content": "Migrate this Express.js app to Fastify."}]
)

High effort suits architecture decisions and security-sensitive analysis. Medium effort handles most code generation at significantly lower latency and cost.

Vision: Screenshot-to-Code in Production

Fable 5's vision capabilities are production-grade. It can rebuild a web application's source code from a screenshot alone, extract precise numbers from dense scientific figures, and validate UI output by comparing rendered screenshots against design specifications.

The Pokémon FireRed result illustrates the spatial reasoning at work: Fable 5 completed the entire game using only raw game screenshots — no maps, no text guides. That same reasoning applies to navigating complex UIs: identifying components, inferring state, and generating appropriate code or interactions.

For teams building design-to-code pipelines, Fable 5 makes screenshot-based validation a viable production step rather than a manual QA afterthought.

Claude Code Integration

In Claude Code, select claude-fable-5 from the model picker. Multi-day autonomous sessions are supported natively — the model persists context across sessions using file-based memory in your working directory.

For MENA teams managing API cost against project budgets: at $10 input / $50 output, Fable 5 is priced below Mythos Preview while delivering equivalent capability for most tasks. For high-frequency, lower-complexity calls, Claude Sonnet 4.6 ($3/$15) or Haiku 4.5 ($1/$5) remain better choices. Reserve Fable 5 for tasks where SWE-bench-class reasoning is what you're paying for — architecture decisions, large migrations, multi-agent orchestration.

Safety Architecture

Fable 5 uses a classifier-based fallback system. Requests touching three sensitive domains trigger automatic fallback to Claude Opus 4.8:

1. Cybersecurity — exploitation assistance and agentic hacking tasks
2. Biology/Chemistry — dual-use research with bioweapons potential
3. Distillation — extracting model capabilities via synthetic data generation

This happens in fewer than 5% of sessions. For the remaining 95%, Fable 5 performs identically to Mythos 5. External red-team testing exceeded 1,000 hours with no universal jailbreaks found.

For enterprise deployments: Mythos-class traffic carries a 30-day retention requirement with strict privacy controls. Review your data processing agreements before routing sensitive codebases through the API.

Pricing and Platform Summary

What Comes Next

Claude Mythos 5 (claude-mythos-5) sits above Fable 5, available only through Project Glasswing to approved customers. The broader trajectory is clear: Anthropic is productizing Mythos-class capability for general developer use.

Build your agentic pipelines around Fable 5 now. The benchmark lead is real, the API is stable, and Ultra Code with Dynamic Workflows addresses the failure modes that made earlier autonomous coding agents unreliable at scale. The 80.3% SWE-Bench Pro score isn't a ceiling — it's a starting point for what becomes possible when you design systems around it.