Fable Brain

Claude Console and API Platform

5 min read

Claude Console and API Platform

Claude Fable 5 is “generally available on the Claude API, Claude Platform on AWS, Amazon Bedrock” plus Google Cloud and Microsoft Foundry (https://platform.claude.com/docs/en/about-claude/models/introducing-claude-fable-5-and-claude-mythos-5), priced at 50 per million output tokens, with pinned snapshot model IDs, refusal fallback to Opus 4.8, and mandatory 30-day retention.

What it is

The developer access layer for the Claude model family: the first-party Claude API plus cloud distribution. Fable 5 ships on Amazon Bedrock in the US East (N. Virginia) and Europe (Stockholm) regions, on Google Cloud, and on Microsoft Foundry. Claude Mythos 5 is not on this self-serve path at all; it is limited to approved Project Glasswing customers via Anthropic, AWS, or Google Cloud account teams.

How it works

  • Model IDs. Current IDs are claude-fable-5, claude-mythos-5, claude-opus-4-8, claude-sonnet-5, and claude-haiku-4-5-20251001 (alias claude-haiku-4-5). Every Claude model ID is a pinned snapshot; dateless IDs from the 4.6 generation onward are also pinned snapshots, not evergreen pointers.
  • Specs. Fable 5 and Mythos 5 share a 1M-token context window by default, up to 128k output tokens per request, a January 2026 knowledge cutoff, and the Opus 4.7 tokenizer, which produces roughly 30 percent more tokens than pre-4.7 models for the same text.
  • Pricing. Fable 5 and Mythos 5: “50 per million output tokens” (https://www.anthropic.com/news/claude-fable-5-mythos-5), less than half of Claude Mythos Preview, with a 90 percent prompt-caching input discount and US-only inference at 1.1x pricing. Below them: Opus 4.8 at 25 (1M context), Sonnet 5 at 15 (1M context), Haiku 4.5 at 5 (200k context).
  • Refusal semantics. “A refusal is a successful HTTP 200 response” (https://platform.claude.com/docs/en/build-with-claude/refusals-and-fallback) with stop_reason refusal and a stop_details object naming the category (cyber, bio, frontier_llm, reasoning_extraction). A refusal arriving before any output is not billed and does not count against rate limits.
  • Fallback paths. Refused Fable 5 requests reach Claude Opus 4.8 three ways: the server-side fallbacks parameter (beta header server-side-fallback-2026-06-01), SDK refusal-fallback middleware (TypeScript, Python, Go, Java, C#), or manual retry with fallback credit; rerouted requests bill at the serving model’s prices.
  • Data handling. Fable 5 and Mythos 5 are Covered Models: they “carry 30-day data retention and are not available under zero data retention” (https://platform.claude.com/docs/en/about-claude/models/introducing-claude-fable-5-and-claude-mythos-5), with retained data used for safety purposes, not training.
  • Thinking. Adaptive thinking is always on; depth is controlled with the effort parameter, budget_tokens returns a 400 error, and raw chain of thought is never returned (thinking.display yields a summary or empty field).
  • Availability history. GA on June 9, 2026; all access suspended under US export controls from June 12; Anthropic reports the controls lifted June 30 with Fable 5 restored globally July 1 and Mythos 5 restored for a set of approved US organizations.

Key insight

Fable 5’s API contract makes safety a first-class response state, not an error: a classifier decline arrives as a normal 200 with stop_reason refusal, is free if it fires before output, and has documented fallback machinery. Clients that only handle HTTP errors will misread designed behavior as an outage.

Best practice

  • Pin behavior expectations to the snapshot ID you deploy; there is no evergreen pointer to silently upgrade you. EVIDENCE-BASED
  • Wire a fallback path before production: server-side fallbacks parameter or SDK middleware, so classifier refusals degrade to Opus 4.8 instead of failing. EVIDENCE-BASED
  • Budget with the new tokenizer in mind; identical text costs roughly 30 percent more tokens than on pre-4.7 models. EVIDENCE-BASED
  • Exploit the 90 percent prompt-caching input discount for repeated system prompts and corpora. EVIDENCE-BASED
  • Check zero-data-retention requirements first; Covered Model status makes Fable 5 ineligible, which disqualifies some compliance regimes. EVIDENCE-BASED
  • Confirm regional availability per cloud; Bedrock coverage is N. Virginia and Stockholm, not every region. EVIDENCE-BASED
  • Consider US-only inference deliberately; the 1.1x price premium buys a data-locality property, not capability. EVIDENCE-BASED

Pitfalls

  • Treating a refusal as an HTTP error; it is a 200 with stop_reason refusal, and error-path-only handling misses it.
  • Billing surprises on fallback: rerouted requests are billed at the serving model’s prices, not Fable 5’s.
  • Sending budget_tokens or prefilled assistant turns; both return 400 errors on this model generation.
  • Assuming availability is stable: all access was suspended June 12 to June 30, 2026 under US export controls, and restoration timing is reported inconsistently across Anthropic and press.
  • Requesting Mythos 5 through the console; “Access is invitation-only and there is no self-serve sign-up.” (https://platform.claude.com/docs/en/about-claude/models/introducing-claude-fable-5-and-claude-mythos-5).

Sources

Next actions

  • Add SDK middleware language coverage notes once a fallback implementation is tested here.
  • Re-verify Bedrock region list at the next refresh; regional rollouts expand quickly.
  • Record observed stop_details payloads from a real refusal to ground the semantics claim.
  • Compare cached versus uncached input costs on this vault’s recurring corpus prompts.

Graph View

Backlinks