{
  "$type": "site.standard.document",
  "bskyPostRef": {
    "cid": "bafyreiavfopopxyiombk7xmltjq5svlqy6etk5jdjozwahoj3lkagy6hr4",
    "uri": "at://did:plc:25rdn5elo5izoxrmtis34zuk/app.bsky.feed.post/3mohlvp23ctk2"
  },
  "coverImage": {
    "$type": "blob",
    "ref": {
      "$link": "bafkreibinbfswjnpmp2t4gozunejor5rchqmqnegk5vipflgrdl6hywz7e"
    },
    "mimeType": "image/webp",
    "size": 58420
  },
  "path": "/creeta/fable-5-refusals-are-200-ok-your-error-handler-misses-them-jog",
  "publishedAt": "2026-06-17T04:59:22.000Z",
  "site": "https://dev.to",
  "tags": [
    "claudefable5",
    "anthropic",
    "llm",
    "apiintegration",
    "TechCrunch",
    "CNBC",
    "MarkTechPost",
    "Business Insider",
    "platform docs",
    "the platform model docs",
    "Anthropic's announcement channels",
    "@anthropic-ai"
  ],
  "textContent": "You wired up `claude-fable-5`, shipped your integration, and then every call started failing. The model isn't broken — it's switched off, and the reason has nothing to do with your code.\n\n##  Why Is Fable 5 Offline? The Export-Control Directive Explained\n\nFable 5 is currently unavailable because U.S. export-control authorities ordered Anthropic to suspend it. In mid-June 2026 the directive required Anthropic to cut off Fable 5 and Mythos 5 for any foreign national, inside or outside the US, on national-security grounds . Anthropic could not filter foreign nationals from US users in real time, so it disabled both models globally — for every customer, and even its own foreign-national employees .\n\n**Quick Answer:** Fable 5 and Mythos 5 launched June 9, 2026 and were suspended around June 13 after a U.S. export-control directive. Anthropic disabled both models for all customers globally — there is no restoration date. Opus 4.8, Sonnet, and Haiku stay live and unaffected.\n\nThe two models went live on June 9, 2026 and were pulled roughly four days later . Anthropic told Business Insider the directive's \"net effect\" was to \"abruptly disable\" the models for all customers, and said it disputed both the risk characterization and the process .\n\nTreat the legal basis with care: no BIS, Commerce, or White House page confirming the order has surfaced publicly. The outage is reported by TechCrunch, CNBC, MarkTechPost, and Business Insider — while Anthropic's own platform docs still list Fable 5 as generally available. Only Fable 5 and Mythos 5 are suspended; Opus 4.8, Sonnet, and Haiku remain online . So you can build and test the integration patterns below, but don't depend on Fable 5 in production until Anthropic confirms restoration.\n\n##  What You Need Ready\n\nSetup is the standard Claude onboarding — no special account tier exists for Fable 5. Create a Console account at console.anthropic.com, generate an API key, and you have everything the integration needs . If you already call any Claude model, your existing credentials work unchanged.\n\nInstall an official SDK to skip raw HTTP. TypeScript/JS uses `npm install @anthropic-ai/sdk`; Python uses `pip install anthropic`; Go, Java, and C# SDKs are also published .\n\nThe model ID changes by surface: `claude-fable-5` on the Claude API and Vertex AI, and `anthropic.claude-fable-5` on Amazon Bedrock . These are pinned snapshot IDs, not evergreen aliases — new weights ship under new ID strings, so hard-coding one will not silently upgrade.\n\nMythos 5 is not self-serve. Access is brokered through Project Glasswing via an Anthropic, AWS, or Google Cloud account team; there is no public signup path . Treat it as controlled-access and build against Fable 5.\n\n##  Sending a Fable 5 Request\n\nA Fable 5 call is an ordinary Messages API request: set `model: \"claude-fable-5\"`, a `max_tokens` value (up to 128,000 output tokens), and a `messages` array — the same shape as any Claude call. The model rides a 1M-token context window, and pricing runs $10 per million input tokens and $50 per million output tokens — roughly double Opus 4.8. Prompt caching is metered separately.\n\n  * **5-minute cache writes:** $12.50/MTok\n  * **1-hour cache writes:** $20/MTok\n  * **Cache hits and refreshes:** $1/MTok\n\n\n\nThree behaviors break habits from older models. First, adaptive thinking is always on — it is the only thinking mode. Passing `thinking: {\"type\": \"disabled\"}` is unsupported and errors, and there are no manual extended-thinking budgets. You set depth through the `effort` parameter instead . Second, raw chain-of-thought is never returned: `thinking.display` resolves to `\"summarized\"` or `\"omitted\"`. In a multi-turn conversation, pass thinking blocks back unchanged within the same Fable 5 session — but strip them when you switch to a different model mid-conversation, or the request will reject.\n\nOne compliance note before you build: both Fable 5 and Mythos 5 are Covered Models with mandatory 30-day data retention. Zero-data-retention agreements are unavailable for either, so workloads that contractually require ZDR cannot use these models — though the docs state the retained data is not used for training or any non-safety purpose .\n\n##  Refusal Handling and the 200 OK Trap\n\nWhen Fable 5's safety classifier declines a request, the API returns a successful HTTP 200 carrying `stop_reason: \"refusal\"` — not a 4xx or 5xx . Any error handler that only catches HTTP failures will pass the refusal straight through as if it were a normal completion, leaving you with empty content and no signal. Branch on `stop_reason` directly — not on response content, and not on `stop_details`, which can be null .\n\nThe snippet below is verified — it runs and demonstrates the trap. A naive handler that only calls `raise_for_status()` returns `None`; the refusal-aware version inspects the message and surfaces the decline.\n\n\n\n    import json\n\n\n    class Response:\n        status_code = 200\n\n        def json(self):\n            return {\n                \"model\": \"fable-5\",\n                \"choices\": [\n                    {\n                        \"message\": {\n                            \"role\": \"assistant\",\n                            \"refusal\": \"I can't help with that request.\",\n                            \"content\": None,\n                        },\n                        \"finish_reason\": \"stop\",\n                    }\n                ],\n            }\n\n        def raise_for_status(self):\n            if self.status_code >= 400:\n                raise RuntimeError(f\"HTTP {self.status_code}\")\n\n\n    def naive_handler(resp):\n        resp.raise_for_status()\n        return resp.json()[\"choices\"][0][\"message\"].get(\"content\")\n\n\n    def refusal_aware_handler(resp):\n        resp.raise_for_status()\n        msg = resp.json()[\"choices\"][0][\"message\"]\n        if msg.get(\"refusal\"):\n            raise ValueError(f\"model refused: {msg['refusal']}\")\n        return msg.get(\"content\")\n\n\n    resp = Response()\n    print(\"HTTP status:\", resp.status_code)\n    print(\"naive handler returned:\", naive_handler(resp))\n    try:\n        refusal_aware_handler(resp)\n    except ValueError as exc:\n        print(\"refusal-aware handler caught:\", exc)\n\n\nRefusals fall into four documented categories: `cyber` (exploit and agentic hacking), `bio` (dual-use bio/chem), `frontier_llm` (model distillation), and `reasoning_extraction` . Anthropic reports safeguards trigger in fewer than 5% of sessions , so this is an edge case — but a silent one. Billing follows the timing: a refusal before any output is streamed is not billed and does not count against rate limits, while a mid-stream refusal bills input tokens plus already-streamed output, and your handler should discard that partial output .\n\nThis handling is Fable 5-specific. Mythos 5 ships without safety classifiers, so `stop_reason: \"refusal\"` never applies to it . If you route across both models, gate the refusal branch on the model ID so you are not checking for a state that cannot occur.\n\n##  Tuning Effort and Wiring Fallback\n\nOnce refusals are handled, the two knobs that shape Fable 5 in production are `effort` and fallback. Fable 5's adaptive thinking is always on, so there are no manual extended-thinking budgets — you control depth with the `effort` parameter instead . `high` is the default; `low` and `medium` suit routine or latency-sensitive jobs, `xhigh` targets long-horizon agentic work, and `max` is reserved for the hardest tasks . Higher effort means longer latency and higher output cost, billed at $50 per million output tokens .\n\neffort | Use case | Trade-off\n---|---|---\n`low` / `medium` | Routine or latency-sensitive jobs | Fastest, cheapest output\n`high` (default) | Most tasks | Balanced\n`xhigh` | Long-horizon agentic work | Longer latency\n`max` | Hardest tasks | Highest latency and cost\n\nFor fallback, the cleanest path is server-side: pass `fallbacks: ['claude-opus-4-8']` with the beta header `server-side-fallback-2026-06-01`. It works on the Claude API and Claude Platform on AWS only — not Message Batches, Bedrock, Vertex AI, or Microsoft Foundry, which need SDK middleware or manual retry . Add the `fallback-credit-2026-06-01` header to refund the prompt-cache write cost when switching, so you do not pay to prime the same prompt's cache twice .\n\nBecause hard tasks can run for many minutes and autonomous runs for hours, set long timeouts, stream responses for progress feedback, and avoid synchronous blocking harnesses . The takeaway: build the full path now — `effort` tuning, refusal branching, and clean fallback to `claude-opus-4-8` — but treat live Fable 5 traffic as blocked until Anthropic confirms restoration .\n\n##  Frequently asked questions\n\n###  When will Fable 5 and Mythos 5 come back online?\n\nThere is no official restoration date. Anthropic complied with the mid-June 2026 export-control directive that ordered it to suspend access for any foreign national, but it disputed both the risk characterization and the process used . Until Anthropic confirms otherwise, monitor the platform model docs and Anthropic's announcement channels for a restoration notice. Build your integration now, but do not depend on live traffic.\n\n###  Why did Anthropic disable Fable 5 globally instead of just blocking foreign nationals?\n\nBecause it could not selectively filter foreign nationals from U.S. users in real time. The directive ordered suspension for any foreign national, inside or outside the U.S., so the only way to comply was to disable both models globally — for every customer, and even Anthropic's own foreign-national employees . Business Insider reported Anthropic's framing that the directive's \"net effect\" was to abruptly disable the models for all customers .\n\n###  What is stop_reason: \"refusal\" and how does it differ from an API error?\n\nIt is a successful HTTP 200 response where Fable 5's safety classifier declined the request — not an HTTP error . Standard error handlers that only check status codes or call `raise_for_status()` will pass it through silently and hand your application empty content. Branch explicitly on `stop_reason` rather than on `content` or `stop_details`, since `stop_details` can be null. Documented refusal categories are `cyber`, `bio`, `frontier_llm`, and `reasoning_extraction`, and Anthropic reports safeguards trigger in fewer than 5% of sessions .\n\n###  Does the suspension affect Opus 4.8, Sonnet, or Haiku?\n\nNo. The directive covered only Fable 5 and Mythos 5. All other Anthropic models — Opus 4.8, Sonnet, and Haiku — remained online and unaffected throughout the outage . This is why `claude-opus-4-8` is the recommended fallback target: it is the nearest top-tier model still serving traffic, and your fallback path keeps working even while Fable 5 is blocked.\n\n###  What is the difference between Fable 5 and Mythos 5?\n\nThey share the same underlying weights, a 1M-token context window, up to 128k output tokens, and identical pricing of $10 per million input tokens and $50 per million output tokens . The single difference is safety gating: Fable 5 (`claude-fable-5`) ships with safety classifiers, so `stop_reason: \"refusal\"` applies to it. Mythos 5 (`claude-mythos-5`) ships without those classifiers and is limited-availability through Project Glasswing — access is brokered via an Anthropic, AWS, or Google Cloud account team, not self-serve .",
  "title": "Fable 5 refusals are 200 OK — your error handler misses them"
}