Claude Fable 5 API: Migrate Your Agents Now (It’s Free Until June 22)

Anthropic released Claude Fable 5 on June 9 — yesterday — and if you’re on Pro, Max, Team, or Enterprise, you’re already paying for the most capable model Anthropic has ever made publicly available. It’s free until June 22. Fable 5 is the first Mythos-class model outside of government use: it scores 80.3% on SWE-bench Pro versus Opus 4.8’s 69.2%, handles a 1-million-token context window, and can run unattended agent loops for days. The migration isn’t just a model ID swap. Here’s what actually changes.

The Numbers Are Not Marginal

Fable 5 leads Opus 4.8 by 11 points on SWE-bench Pro and by 21.7 points over GPT-5.5. On FrontierCode Diamond — Cognition’s hardest production-codebase benchmark — it scores 29.3% against Opus 4.8’s 13.4%. That’s more than double. These aren’t incremental gains dressed up as a major release.

The proof that translates to dollars: Stripe used Fable 5 to migrate a 50-million-line Ruby codebase in a single day. The same job would have taken five engineers over two months — roughly $200,000 in fully-loaded labor cost, compressed into a few thousand dollars in token usage. That math is hard to ignore if you’re running any serious agent workloads.

Five Things to Change in Your Code

If you’re calling Opus 4.8 today, migrating to Fable 5 requires five changes — not one:

1. Model ID: claude-opus-4-8 → claude-fable-5
2. Remove thinking: {type: "disabled"}: Thinking is always adaptive on Fable 5. This parameter will error.
3. Remove budget_tokens: Gone. There’s no direct replacement.
4. Add effort: This is now how you control depth and cost. Options: low, medium, high (default), xhigh.
5. Handle stop_reason: "refusal": This is a successful HTTP 200 response, not an error. Plan for it.

import anthropic

client = anthropic.Anthropic()  # ANTHROPIC_API_KEY from env

message = client.messages.create(
    model="claude-fable-5",
    max_tokens=4096,
    messages=[{"role": "user", "content": "Your prompt here"}]
)

if message.stop_reason == "refusal":
    # Refused before output — not billed. Retry or fallback.
    pass
else:
    print(message.content)

Run pip install -U anthropic first — older SDK versions have stale model registries and won’t recognize claude-fable-5.

Thinking Mode: Adaptive Is the Only Mode

Fable 5 always thinks. You don’t enable or disable it — you control depth with the effort parameter. At high (the default), the model almost always thinks before responding. At low, it skips thinking for simple problems. If you were using budget_tokens to manage costs on Opus 4.8, your new lever is effort level.

Anthropic’s recommendation: start at high for most tasks, use xhigh for the most demanding agentic workloads. The raw chain-of-thought is never returned, but setting thinking.display to "summarized" gives you readable summaries of what the model was working through. See the adaptive thinking documentation for the full parameter reference.

Safeguards: What Actually Gets Blocked

Less than 5% of sessions hit the classifier. The blocked domains are: cybersecurity exploit code, biological pathogen information, chemical synthesis routes, and model distillation attempts. For the vast majority of developers, the safeguards are invisible.

For those building security tooling or anything adjacent to those domains: test your specific prompts before you ship. When the classifier fires, the API returns stop_reason: "refusal" with an HTTP 200 and a note on which classifier triggered. You’re not billed for refused requests that produced no output. The SDK middleware (Python, TypeScript, Go, Java, C#) includes built-in fallback retry to Opus 4.8 if you need it. Simon Willison’s initial review covers the safeguard behavior in detail for security researchers.

It’s Available Everywhere, Today

Fable 5 launched day-one across the Claude API, Amazon Bedrock, Vertex AI, Microsoft Foundry, and GitHub Copilot — no waitlist, no preview tier. On Bedrock the model ID is anthropic.claude-fable-5 with regional variants like us.anthropic.claude-fable-5. Databricks users get access through Unity AI Gateway.

One gotcha: GitHub Copilot Enterprise and Business administrators need to explicitly enable the Claude Fable 5 policy in Copilot settings. It’s off by default. If your team uses Copilot Business or Enterprise and you don’t see Fable 5 as an option, that’s why.

Twelve Days Free, Then $10/$50 per Million Tokens

Fable 5 is included at no extra cost on Pro, Max, Team, and seat-based Enterprise plans through June 22. After that, it moves to usage credits at $10 per million input tokens and $50 per million output — exactly 2x Opus 4.8. For high-volume pipelines doing simple, repetitive calls, Opus 4.8 likely stays the right call. For anything involving long-horizon agents, large-context analysis, or complex codebases, the performance gap justifies the price.

The Anthropic migration guide covers the full set of changes, including edge cases around prompt caching with the new model and rate limit differences. With 12 days of free access still on the clock, there’s no reason not to run your existing workloads against Fable 5 and see what it actually does with your prompts. Read the full announcement for the complete capability breakdown.