# Claude Fable 5 on Bedrock — model selection decision matrix

Use this when a stakeholder asks *"should we move everything to Fable 5?"* The answer is almost always **no** — route by workload shape, governance constraints, and whether you can accept Anthropic's data-retention opt-in.

> Reflects the **June 9, 2026** GA on Amazon Bedrock. Model IDs, regions, and pricing change — confirm in the [model card](https://docs.aws.amazon.com/bedrock/latest/userguide/model-card-anthropic-claude-fable-5.html) before committing.

## 1. Pick the access path first

| You need… | Use | Why |
|-----------|-----|-----|
| Data stays in AWS infra; compose with Guardrails, Knowledge Bases, Agents | **Amazon Bedrock** (`bedrock-runtime` or `bedrock-mantle`) | IAM, PrivateLink, regional residency, existing Bedrock billing tags |
| Anthropic-native experience (projects, Claude Code-style workflows) with AWS billing | **Claude Platform on AWS** | Same models; different console and API surface — see [Claude Platform docs](https://docs.aws.amazon.com/claude-platform/latest/userguide/welcome.html) |
| Unrestricted Mythos-class model (no Fable safeguards) | **Claude Mythos 5** (preview, vetted customers only) | Not GA for general enterprise — do not architect assuming access |

**We recommend Bedrock** for production apps that already run on AWS governance. Use Claude Platform on AWS for teams that want Anthropic's native UX but need consolidated AWS procurement.

## 2. Pick the model tier

| Workload | Lean **Fable 5** | Lean **Claude 4.6 Sonnet / Opus** | Lean **Nova Pro / Lite** |
|----------|------------------|-------------------------------------|--------------------------|
| Multi-hour coding or research agent (async, self-verifying) | Yes | No — prior models lose coherence on long unbroken runs | No |
| Document-heavy vision (nested tables in PDFs, architecture diagrams) | Yes | Sonnet if docs are short; Opus if reasoning depth matters more than vision | No |
| High-volume chat / classification (< 4K context) | No — overkill cost/latency | Haiku-class or Nova Micro | Yes |
| Regulated workload that **cannot** opt into 30-day provider data retention | **No** | Yes (models without `provider_data_share` requirement) | Yes |
| Prompts touching cybersecurity / bio / chem / health domains | Expect **classifier fallback** (may route to Opus 4.8, lower capability) | Sonnet/Opus if you need predictable capability in those domains | N/A |
| Steady production API with strict p99 latency SLO | Cautious — adaptive thinking is always on | Sonnet for balanced latency | Nova for cost |

## 3. Governance gates (hard stops)

Before routing production traffic to Fable 5 on Bedrock, confirm:

- [ ] Legal/security signed off on **Data Retention API** opt-in (`provider_data_share`) — Anthropic requires **30-day input/output retention + human review**. No console UI at launch; API-only.
- [ ] Application handles `stop_reason: "refusal"` as a **primary path** — refusal rates are materially higher than prior Claude models. Prompt-stage refusals are not billed; mid-stream refusals bill partial output.
- [ ] Inference profile uses a **region compliant with data residency** — in-region only in `us-east-1` and `eu-north-1` at launch; other US/EU regions via geo/global profiles.
- [ ] Cost model accounts for **Standard tier only** at launch (no Priority/Flex/Reserved on Fable 5 yet).
- [ ] Long-running jobs have **checkpointing** — Fable 5 can run extended async work; your orchestrator must survive Bedrock timeouts and retries.

## When NOT to adopt Fable 5 yet

- You have not tested refusal handling on your real prompt distribution — a 3% refusal rate on a 50k/day API is 1,500 user-visible failures.
- Your BAA or internal policy forbids third-party model-provider review of inference payloads.
- The workload is a drop-in replacement for Sonnet on short prompts — you will pay frontier-model prices without using frontier-model duration.
- You need Priority or Reserved throughput commitments today — check the model card; Fable 5 may be Standard-only at launch.