Claude Fable 5 API: Access, Pricing, Use Cases

Anthropic released Claude Fable 5 on June 9, 2026. The API model ID is claude-fable-5. I’ve spent the last few days going through the docs and running it on a couple of real workloads, and this piece is the Claude Fable 5 API access, pricing, and use-case overview I wish I’d had on day one.

Two things worth flagging upfront. This is the first publicly available model in what Anthropic calls the Mythos class — pricing is double Claude Opus 4.8, capability is positioned a tier above. And the Claude Fable 5 API behaves differently from earlier Claude models in some specific ways that matter for production integration. The headline differences are worth understanding before changing a single line of code.

What Claude Fable 5 is for builders

The official Anthropic announcement frames Anthropic Fable 5 as the first generally available Mythos-class model, built to sustain long autonomous tasks across software engineering, knowledge work, vision, and scientific research. That’s the marketing line. What follows is what it means in practice.

Mythos-class capability with public safeguards

Fable 5 and Claude Mythos 5 are the same underlying model. The difference is what each will help with. Mythos 5 is restricted to vetted partners through Project Glasswing and keeps the full set of capabilities, including high-end offensive cybersecurity skill. Fable 5 is the public release, which runs safety classifiers that reroute requests in cybersecurity, biology, chemistry, and model-capability-extraction domains to Claude Opus 4.8 instead.

Anthropic’s framing is that fewer than 5% of sessions trigger the reroute. Outside those four domains, the two models behave the same way.

API model ID and supported platforms

The API model identifier is claude-fable-5. As of the June 9 launch, Fable 5 is available through the Claude API, Claude Platform on AWS, Amazon Bedrock, Vertex AI, and Microsoft Foundry. Claude Code support requires v2.1.170 or later — older versions don’t show Fable 5 in the model picker.

Why this matters for long-horizon agent workloads

The capability gap shows up most clearly on long, multi-step, agentic work. Anthropic’s positioning — and early benchmark coverage — points to noticeably stronger sustained autonomy than Opus 4.8 on tasks measured in hours or days rather than minutes. For builders running multi-stage agent harnesses, that compounds. For builders running one-shot chat, it largely doesn’t. Pricing reflects this gap.

API access and availability

Five paths to reach the model at launch. Each has its own enablement quirks, but the Claude Fable 5 API surface is the same across them.

Claude API and Claude Platform on AWS

Direct access through api.anthropic.com is straightforward — point your existing Claude API SDK at model ID claude-fable-5, and the request lands on the model. The same model ID works on Claude Platform on AWS through Anthropic’s deployment there. The API integration notes Anthropic published alongside the launch walk through the behavioral differences worth knowing — primarily around the refusal stop reason and adaptive thinking parameters.

Amazon Bedrock, Vertex AI, and Microsoft Foundry

For teams already on a cloud provider, Anthropic Fable 5 is available natively on all three major hyperscaler platforms. Bedrock requires a one-time model enablement flow. Vertex AI needs Advanced AI Safety Addendum consent and a project-level enable per region. Foundry follows Microsoft’s standard partner-model onboarding. The wrapper details differ. The underlying model is the same.

Worth checking before assuming: cloud-platform availability of a specific Claude alias (e.g., what opus resolves to on Bedrock vs the Anthropic API) can lag the direct API by a generation. Pinning to the explicit claude-fable-5 string avoids surprises.

Claude.ai access vs API access

As of publication, Fable 5 is included free on Pro, Max, Team, and seat-based Enterprise plans during an introductory window from June 9 through June 22, 2026. After that, consumer plan access shifts to usage credits. The API is the path for builders regardless of plan tier — subscription access is for end-user chat, not for production integration. Verify the current state against Anthropic’s official documentation if you’re reading this later. Introductory pricing windows close quickly.

Covered Model data retention requirements

This one matters and is easy to miss. Both Fable 5 and Mythos 5 are designated Covered Models, which means 30-day data retention applies and zero data retention (ZDR) is not available. If a contract or compliance posture requires ZDR, Fable 5 is not an option for those workloads — you stay on Opus 4.8 or earlier, where ZDR remains supported. Please refer to Anthropic’s official documentation for the current state of data retention policies.

Pricing and cost controls

Cost control on a model this expensive isn’t optional. The model pricing tells one story, but three numbers decide the actual bill: the headline Claude Fable 5 API pricing, prompt caching configuration, and max_tokens discipline. Each compounds.

Official token pricing and why output cost matters

At launch, the published API pricing is $10 per million input tokens and $50 per million output tokens. Exactly double Claude Opus 4.8 on both sides. Worth dwelling on the second number: at $50 per million output, a single 30K-token response — which Fable 5 will happily produce on agentic tasks if uncapped — costs $1.50 in output alone. Long agent sessions accumulate output fast. Capping max_tokens on every API call is the cheapest cost-control intervention available.

1M context, prompt caching, and US-only inference notes

The full 1M-token context window is included at standard pricing — no surcharge for long context. Prompt caching applies the same way it does on other Claude models, with cache hits at $1 per million tokens (90% off the standard input rate). For agent harnesses that send the same system prompt and tool definitions every turn, that’s the difference between a routine bill and a runaway one.

US-only inference is available with a 1.1x multiplier across all token categories for data-residency needs. Default routing is global.

Effort, adaptive thinking, and task budgets

Fable 5 uses adaptive thinking — the model decides its own reasoning depth per request. Extended thinking budgets from Opus 4.8 don’t apply. Instead, an effort parameter trades thoroughness against token spend. The thinking output itself is summarized rather than full. For builders porting prompts from Opus 4.8, this is the parameter change most likely to need attention.

When not to use Fable 5 for lightweight tasks

Fable 5 at 2x Opus model pricing is not the default. For simple chat, classification, format conversion, or any task where Opus 4.8 or Sonnet 4.6 already gets you to acceptable quality, those models are the correct choice. Reserve Fable 5 for workloads where the capability gap actually pays for itself — long agentic sessions, complex codebase migrations, multi-stage research tasks. Routing by task complexity is the cost-control lever that matters most over a month.

Refusals, fallback, and safety filters

Sensitive domains that may trigger classifiers

The safety classifiers Fable 5 runs target offensive cybersecurity work (building exploits, malware, attack tooling), biological weapon design, chemical weapon design, and model-capability-extraction techniques. Defensive security research, general bug-finding outside those specific categories, and most cybersecurity-adjacent professional work runs through normally. Edge cases exist, and what counts as “edge” gets refined as Anthropic gathers data.

Handling `stop_reason: "refusal"` in production

When the classifiers trigger, the API response carries stop_reason: "refusal" rather than a normal completion. This is the new field integration code needs to handle. For independent context on the safety architecture behind the refusal behavior, TechCrunch’s launch coverage summarizes Anthropic’s public reasoning for the split between Fable and Mythos.

Keep fallback implementation details for the dedicated fallback guide

I’m not walking through fallback patterns here. Detecting stop_reason: "refusal", retrying on Opus 4.8 versus another model, billing implications for rerouted calls, structured error handling — those deserve their own piece. For the integration overview, the relevant facts are: the classifier exists, it’s narrow, the response shape is documented, and rerouted requests are not billed at Fable rates. The deeper integration patterns belong in a dedicated fallback guide.

Best-fit builder use cases

Codebase migration and agentic coding

The clearest use case I’ve seen is long-horizon coding work — multi-day migrations, multi-stage refactoring, agent harnesses that sustain context and self-verify across many turns. Claude Code Fable (the Fable 5 integration documented in Anthropic’s Claude Code model config guide) targets exactly this scenario: tasks “larger than a single sitting.” Claude Code Fable shines on the multi-day end of that spectrum, not the one-off end. For single-file edits or quick scripts, Sonnet 4.6 still wins on cost and latency.

Complex reasoning and knowledge work

Long research tasks, document analysis at the 500K+ token scale, multi-source synthesis. The 1M context window matters here, but the binding constraint is usually reasoning depth on ambiguous inputs rather than raw context size. This is where the capability gap over Opus 4.8 shows up most clearly, in my observation so far.

Vision-heavy app prototyping

Anthropic positions Fable 5 as state-of-the-art on vision — reading scientific figures, rebuilding web app source code from screenshots, navigating UIs with minimal scaffolding. For builders prototyping vision-first apps, that’s a meaningful starting point. To be verified for any specific workload — vision benchmark generalization is famously uneven, and your screenshots are not the benchmark.

FAQ

Is Claude Fable 5 available through the API?

Yes, since June 9, 2026. The Claude Fable 5 API model ID is claude-fable-5, accessible through the Claude API directly, Claude Platform on AWS, Amazon Bedrock, Vertex AI, and Microsoft Foundry. Claude Code requires v2.1.170 or later.

Which platforms host Claude Fable 5?

Five at launch: the Claude API, Claude Platform on AWS, Amazon Bedrock, Vertex AI, and Microsoft Foundry. Each cloud provider has its own enablement flow — Bedrock needs a one-time model enablement, Vertex AI requires Advanced AI Safety Addendum consent, Foundry follows Microsoft’s partner-model onboarding. The underlying model is the same across all of them. Please refer to the official documentation for the current state of platform-specific access requirements.

How is Claude Fable 5 different from Claude Mythos 5?

Same underlying Mythos-class model. Fable 5 is the public release with safety classifiers active in cybersecurity, biology, chemistry, and capability-extraction domains. Mythos 5 is restricted to approved partners through Project Glasswing and runs without those classifiers. The Fable-vs-Mythos comparison has more dimensions worth working through, and a dedicated comparison piece covers that in depth.

What happens if Fable 5 refuses a request?

The API response carries stop_reason: "refusal" instead of a normal completion. Anthropic’s framing is that this affects fewer than 5% of sessions, focused on the four sensitive domains above. Builders need to handle the new stop reason in production code — detecting it, optionally retrying on Opus 4.8 or another model. Please refer to Anthropic’s latest documentation for the current refusal and fallback specifications.

Conclusion

Claude Fable 5 is the first Mythos-class model anyone can call directly, and the Claude Fable 5 API integration is straightforward — one model ID, five hosting platforms, public token pricing. The decisions that matter aren’t about access; they’re about routing. When to send a task to Fable 5 versus a cheaper model. When to absorb the 2x cost. When to expect a refusal and how to fall back. None of those are answered by reading an announcement.

Run it yourself on a real workload. That’ll tell you more than anything I say.

More to come on the Fable-vs-Mythos comparison and on fallback patterns specifically.

Previous posts: