Claude Fable 5: The Complete Developer Guide (Benchmarks, API & the June 23 Billing Shift)

Stripe took a 50-million-line Ruby codebase, ran a full migration with Claude Fable 5 in a single day, and estimated it would have taken a senior engineering team more than two months by hand. That one data point tells you more about what changed on June 9, 2026 than any benchmark chart.

Claude Fable 5 is Anthropic's first publicly available Mythos-class model — a capability tier above Opus that didn't exist six months ago. If you're on a Pro, Max, Team, or Enterprise plan, you've been using it for free since June 9. That window closes June 22. Starting June 23, continued use draws down usage credits at standard API rates.

Here's everything you need to know: what it is, how fast it actually is on real tasks, how to call it from code, and when to route to it instead of Opus 4.8.

What Is the Mythos Class — and Why Does It Matter?

Anthropic's model hierarchy now has four public tiers: Haiku, Sonnet, Opus, and Mythos. Fable 5 is the first Mythos model available to everyone. Claude Mythos 5 (a more capable sibling) remains restricted to trusted-access research partners for now.

The Mythos class is built for a specific use case: long-horizon agentic work — tasks that take hours or days, require dozens of tool calls, involve delegating to sub-agents, and need the model to validate its own outputs across a multi-step pipeline. Standard Opus-class models handle complex reasoning exceptionally well but were designed around single-session, human-in-the-loop workflows. Fable 5 was designed to work without constant human checkpoints.

This matters practically because:

• It sustains coherent reasoning across 1 million tokens of context (up to 128,000 output tokens per request)
• It can delegate to specialist sub-agents and reconcile results on a shared filesystem
• It self-validates before returning results, reducing the hallucination rate on complex multi-step chains
• Its reasoning advantage over Opus 4.8 grows with task complexity — on simple tasks the gap is small; on hard agentic tasks it's dramatic

Benchmark Results: Where Fable 5 Pulls Away

Raw benchmark numbers matter less than which benchmarks you choose. The two that matter most for developers are SWE-Bench Pro and FrontierCode Diamond.

SWE-Bench Pro (Real GitHub Issue Resolution)

SWE-Bench Pro tests whether a model can read a real GitHub issue, navigate an unfamiliar production codebase, and produce a working patch. It's the closest public proxy for autonomous coding agent performance.

The 11-point gap between Fable 5 and Opus 4.8 is larger than the gap between Opus 4.8 and Google's Gemini 3.1 Pro. And importantly, Fable 5's SWE-Bench lead widens further on multi-file patches and cross-module refactors — exactly the class of tasks Claude Code agents handle in production.

FrontierCode Diamond (Production-Grade Difficulty)

This benchmark explicitly filters out tasks that can be solved by pattern-matching to training data. It targets problems that require genuine reasoning under production constraints.

Fable 5 more than doubles Opus 4.8 on this benchmark. A 29% score may sound low — but the problems are intentionally unsolvable by anything that can't actually reason, which means the gap is the signal.

Claude Fable 5 API: Quickstart

The model string is claude-fable-5. You call it exactly like any Claude model:

pythonimport anthropic

client = anthropic.Anthropic()

message = client.messages.create(
    model="claude-fable-5",
    max_tokens=8192,
    messages=[
        {
            "role": "user",
            "content": "Analyze this 500-page technical spec and identify every compliance gap against ISO 27001."
        }
    ]
)

print(message.content[0].text)

For agentic workflows with tool use:

pythonimport anthropic

client = anthropic.Anthropic()

tools = [
    {
        "name": "read_file",
        "description": "Read a file from the codebase",
        "input_schema": {
            "type": "object",
            "properties": {
                "path": {"type": "string", "description": "File path to read"}
            },
            "required": ["path"]
        }
    }
]

response = client.messages.create(
    model="claude-fable-5",
    max_tokens=16384,
    tools=tools,
    messages=[
        {
            "role": "user",
            "content": "Audit the authentication module for security vulnerabilities."
        }
    ]
)

The 1M context window means you can pass enormous codebases, lengthy documentation, or full conversation histories in a single call. Fable 5 is designed to make effective use of that context rather than ignoring tokens past a certain point.

Pricing and the June 23 Billing Change

This is what most developers need to act on right now.

• Input tokens: $10 per million
• Output tokens: $50 per million
• That's 2× the price of Opus 4.8 ($5/$25 per million)

• June 9–22: Fable 5 access is free — no usage credit deductions
• June 23 onward: Usage credits are required at standard API rates

If you're a heavy user on Pro or Max who has been leaning on Fable 5 for complex tasks, plan for the billing change. The practical implication: use Fable 5 for tasks where the quality difference genuinely matters (long-form analysis, multi-step code migrations, complex document review), and route everyday tasks to Sonnet or Opus 4.8 to manage credit consumption.

One important note for privacy-sensitive deployments: Fable 5 currently requires a 30-day data retention period. Opus 4.8 supports zero data retention. If your use case requires no-retention processing (healthcare, legal, certain financial applications), Opus 4.8 remains the right choice until Anthropic releases a Fable 5 zero-retention tier.

Claude Fable 5 vs Claude Opus 4.8: When to Use Each

The right mental model: Opus 4.8 is your everyday workhorse; Fable 5 is what you reach for when you'd otherwise book a senior engineer for half a day.

In production, smart routing pays off fast. Treat Fable 5 as the top escalation tier: try Opus 4.8 first, escalate to Fable 5 only when the task complexity warrants it. A/B test with real prompts to measure where the acceptance rate, retry count, and latency differences actually justify the 2× price increase.

What Fable 5 Means for the Claude Ecosystem

Fable 5 signals a shift in what Anthropic considers a production-ready model. The Mythos class is explicitly designed for AI systems that operate over days, not sessions. That means:

Key Takeaways

• Claude Fable 5 (claude-fable-5) is Anthropic's first Mythos-class model — one tier above Opus, built for multi-day autonomous agentic work
• 80.3% SWE-Bench Pro (vs 69.2% for Opus 4.8) — a gap that grows on harder coding tasks
• 1M context, 128K output — designed for large codebase analysis, long documents, and multi-step pipelines
• Pricing: $10/$50 per million input/output tokens — 2× Opus 4.8, warranted for complex tasks
• Free on Pro/Max/Team/Enterprise through June 22 — billing changes June 23, plan accordingly
• Requires 30-day data retention — Opus 4.8 remains the choice for zero-retention requirements
• Route Fable 5 to high-complexity tasks; keep Opus 4.8 as the default to manage costs

Next Steps

If you're building production systems with Claude, the strategic question is no longer "which Claude model should I use?" but "how do I build intelligent routing logic that escalates to Fable 5 only when the task warrants it?"

That's a model architecture question — and it's exactly what the Claude Certified Architect (CCA-F) certification tests. Our CCA exam prep guide walks through model selection frameworks, multi-agent design patterns, and the kind of architectural reasoning the exam covers.

You can also explore our Claude multi-agent orchestration tutorial for practical examples of building Fable 5 lead agents with Opus 4.8 subagents — the cost-efficient pattern most production teams will land on.