Skip to content

Claude Fable 5 — Knowledge Base

Research note

This page is background research kept for the project's own reference. It is not user documentation, it is not kept up to date with every release, and model or product details in it may be out of date. For how anti-hall works today, start at the home page.

Compiled 2026-06-10 from 14 sources (researcher agent draft; load-bearing claims under independent spot-check — see Verification status at the end). All claims carry inline source references [n] mapped to the numbered list at the end.

[UPDATE 2026-08-22 — re-verified against the authoritative reference; this doc's core Fable 5 claims hold up better than the other model KBs. Confirmed still accurate: claude-fable-5 id, Fable 5 = current flagship, $10/$50 pricing, 1M context / 128k output, five-tier effort scale (default high). Two things are now stale downstream of this doc's compile date: (1) Opus 4.8 (referenced throughout §1/§3/§5/§6/§7 as the comparison/fallback model) is now DEPRECATED / legacy — succeeded by Opus 5 (claude-opus-5, same $5/$25, 1M/128k, shipped 2026-06-09); the §7 migration guide title ("Opus 4.8 → Fable 5") is still the correct migration path per the vendor's own migration guide, just note the source model is now legacy, not current. (2) Sonnet 4.6 (§1 tier hierarchy, §3 pricing) was accurate at this doc's 2026-06-10 compile date but has since been superseded by Sonnet 5 (claude-sonnet-5, released 2026-06-30, $2/$10 — the Sept-1 increase to $3/$15 that other KBs cite was cancelled) — not an error in this doc, a generation that shipped after it was written. Cache pricing (write 5-min 1.25×/write 1-hr 2×/read 0.1× base) is confirmed to apply across all current models including Fable 5. Left as stacked annotation per this KB's own draft/spot-check convention, not rewritten.]**

[UPDATE 2026-09-03 — Fable point release: claude-fable-5-1 is now current, superseding claude-fable-5. claude-fable-5-1 was released 2026-09-01. This is a point release of Fable, not a new generation — Opus 5 / Sonnet 5 / Haiku 4.5 lineage above is unaffected by this update. Pricing, context window, and effort-scale specifics for claude-fable-5-1 are NOT independently verified in this KB — do not assume they are unchanged from Fable 5's $10/$50, 1M/128k without checking the vendor docs. Every claude-fable-5 ID and "current flagship" claim throughout this doc (§1 identity table, §1 tier hierarchy, §4 availability, §7 migration title) refers to the now-superseded prior point release; left as historical per this KB's own stacked-annotation convention. Route by tier token (model:"fable") in all live spawn paths — per MODEL-POLICY.md, the token resolves to the newest Fable build at call time, so no routing code needs a version bump for this change.]**

1. Identity

Model IDs

Surface ID
Claude API / Claude Platform on AWS claude-fable-5
Amazon Bedrock anthropic.claude-fable-5
Bedrock geo-routing us.anthropic.claude-fable-5 / eu.anthropic.claude-fable-5
Vertex AI claude-fable-5
Microsoft Foundry / Azure AI claude-fable-5

Source: [1][2][7]

Release date: June 9, 2026. [1][2]

Naming: "Fable is from the Latin fabula, 'that which is told,' akin to the Greek mythos. The safeguards are what distinguish the two models." [3] Fable 5 is the safeguarded release of the Mythos-class model; Mythos 5 is the access-gated version without general-use classifiers. [1][3]

Tier hierarchy (top→bottom): 1. Claude Mythos 5 (claude-mythos-5) — same weights, classifiers lifted, invitation-only (Project Glasswing) [1][2] 2. Claude Fable 5 (claude-fable-5) — most capable generally-available model [1][2] 3. Claude Opus 4.8 (claude-opus-4-8) [2] 4. Claude Sonnet 4.6 (claude-sonnet-4-6) [2] 5. Claude Haiku 4.5 (claude-haiku-4-5) [2]

[UPDATE 2026-08-22] Re-anchored against the verified current lineup: rows 1-2 unchanged (Mythos 5 / Fable 5 still top of tier). Row 3, Opus 4.8, is superseded by Opus 5 (claude-opus-5, deprecated/legacy status for 4.8, same price/context class). Row 4, Sonnet 4.6, is superseded by Sonnet 5 (claude-sonnet-5, $2/$10, released 2026-06-30, after this doc's compile date). Row 5, Haiku 4.5, unchanged (200k context, still below the 1M-context tier — disqualified as a main-agent downshift target, see KB-model-modes.md §13).

2. Capabilities

Spec Value
Context window 1M tokens (Opus 4.7 tokenizer — same text ≈ 30% more tokens than pre-4.7) [2]
Max output (sync) 128k tokens [1][2]
Max output (Batch) UNVERIFIED for Fable 5 (Opus 4.8 has 300k via output-300k-2026-03-24 beta) [2]
Knowledge cutoff January 2026 [7]

Thinking: Adaptive thinking is the ONLY mode — always on; thinking: {type: "disabled"} is rejected ("not supported" per official docs; exact error code UNVERIFIED — spot-check found no "400" wording), and manual budget_tokens extended thinking is likewise not supported (effort controls depth instead). Raw CoT never returned (thinking.display defaults "omitted"; "summarized" available). Pass thinking blocks back unchanged in multi-turn on the same model. [1][4]

Effort parameter — all five levels supported on Fable 5 [4]: low / medium / high (default) / xhigh (Fable 5, Mythos 5, Opus 4.8/4.7 only) / max (reserve for frontier problems). Official note: low is "suitable for subagents". "Ultracode" is NOT an API effort level — it pairs xhigh with multi-agent permission messaging in Claude Code. [4] At high/xhigh set large max_tokens (hard cap on thinking + response). [4]

Agentic features at launch: tool use, memory tool, task budgets (beta), context editing/tool-result clearing (beta), compaction, vision. [1] AWS model card: "sustained autonomous operation across multi-day tasks — plans across stages, delegates to sub-agents, and self-verifies its work." [7][3]

Bedrock sampling constraints: temperature 1.0/unset; top_p ≥0.99 <1.0/unset; no top_k. [7]

3. Pricing (per MTok)

Model Input Output vs Fable input vs Fable output
Fable 5 $10 $50 — —
Opus 4.8 $5 $25 2× cheaper 2× cheaper
Sonnet 4.6 $3 $15 3.3× cheaper 3.3× cheaper
Haiku 4.5 $1 $5 10× cheaper 10× cheaper

Sources: [1][2]. Fable 5 < half the price of Mythos Preview. [3] Bedrock prompt caching: 90% input discount (1,024-token min, 4 checkpoints, 5-min/1-h TTL). [7] Refusals before any output are not billed; mid-stream refusals billed to the block point; "fallback credit" refunds prompt-cache cost on fallback-model retry. [1]

4. Availability

  • GA 2026-06-09: Claude API, Claude Platform on AWS, Bedrock, Vertex AI, Microsoft Foundry. [1][2]
  • Claude.ai: Free/Pro/Enterprise (Free phased in after 2026-06-22, credits beyond allotment). [5]
  • Claude Code: by model name claude-fable-5; also via "ultracode" mode. [4]
  • GitHub Copilot: Pro+/Max/Business/Enterprise; org policy DISABLED by default (data-retention requirement). [10]
  • Bedrock tiers: Standard only at launch (no Priority/Flex/Reserved). [7]
  • ⚠ Data retention: Fable 5 is a Covered Model — mandatory 30-day retention, NOT available under Zero Data Retention; existing ZDR commitments don't extend to it (classifier operation requirement). Opus 4.8 / Sonnet 4.6 / Haiku 4.5 still support ZDR. [1][8][10]

5. Benchmarks vs Opus 4.8 (launch materials; independent replication limited)

Benchmark Fable 5 Opus 4.8 Source
SWE-Bench Pro 80.3% 69.2% [6]
FrontierCode Diamond 29.3% 13.4% [6]
GDP.pdf vision (no tools) 29.8% 22.5% [6]
Hex core analytics "first to break 90%" — [5]
Hebbia Finance highest scorer — [3]

Effort scaling: FrontierCode Diamond 11.5%→30.9% (low→xhigh). [6] Counterpoint: Andon Labs reported Mythos 5 underperforming some competitors on agentic-business benchmarks. [6]

6. Orchestration guidance (feeds anti-hall model routing)

  • Safety fallback routing: 3 classifiers (cyber, bio/chem, distillation) redirect refused requests to Opus 4.8 when API fallbacks is set; >95% of sessions involve no fallback. [1][3][5]
  • Cost tiering [4][9]: Fable 5 high/xhigh/max = orchestrator + hardest subtasks; Fable low/medium = lower-priority subtasks (still often exceeds xhigh on Opus 4.7); Opus 4.8 xhigh = complex agentic/coding where Fable cost unjustified; Sonnet 4.6 = most implementation; Haiku 4.5 = lookups, classification, navigation, high-volume subagents.
  • Multi-agent: async subagent harnesses (3/5/10 agents) → 2.2×/2.7×/2.7× latency speedups vs single-agent; non-blocking beats blocking on latency AND tokens. [9]

7. Migration notes (Opus 4.8 → Fable 5) [1][4][14]

  1. Remove thinking: {type:"disabled"} (rejected — official docs say "not supported"; exact error code unverified).
  2. Remove manual budget_tokens extended thinking (rejected — official docs say "not supported"; exact error code unverified).
  3. max_tokens caps thinking+response combined — revisit values.
  4. Handle stop_reason: "refusal" (HTTP 200 path, not an error).
  5. Raw thinking never returned.
  6. Strip Fable-5 thinking blocks before replaying history to non-Fable/Mythos models.
  7. Tokenizer ≈ +30% tokens vs pre-4.7 — re-baseline costs. [2][14]
  8. No ZDR (see §4). No Opus 4.8 deprecation announced; Sonnet 4 / Opus 4 retire 2026-06-15, Opus 4.1 retires 2026-08-05 (unrelated). [2]

Sources

  1. Anthropic API Docs — Introducing Claude Fable 5 and Claude Mythos 5 (2026-06-10): https://platform.claude.com/docs/en/about-claude/models/introducing-claude-fable-5-and-claude-mythos-5
  2. Anthropic API Docs — Models Overview (2026-06-10): https://platform.claude.com/docs/en/about-claude/models/overview
  3. Anthropic — Fable 5 / Mythos 5 announcement (2026-06-10): https://www.anthropic.com/news/claude-fable-5-mythos-5
  4. Anthropic API Docs — Effort parameter (2026-06-10): https://platform.claude.com/docs/en/build-with-claude/effort
  5. TechCrunch — Anthropic released Claude Fable 5 (2026-06-10): https://techcrunch.com/2026/06/09/anthropic-released-claude-fable-5-its-most-powerful-model-publicly-days-after-warning-ai-is-getting-too-dangerous/
  6. Vellum.ai — Fable 5 / Mythos 5 benchmarks explained (2026-06-10): https://www.vellum.ai/blog/claude-fable-5-and-mythos-5-benchmarks-explained
  7. AWS Bedrock — Model card: Claude Fable 5 (2026-06-10): https://docs.aws.amazon.com/bedrock/latest/userguide/model-card-anthropic-claude-fable-5.html
  8. Anthropic Help Center — Data retention for Mythos-class models (2026-06-10): https://support.claude.com/en/articles/15425996-data-retention-practices-for-mythos-class-models
  9. OpenRouter — Claude Fable 5 API specs (2026-06-10): https://openrouter.ai/anthropic/claude-fable-5
  10. GitHub Changelog — Fable 5 GA for Copilot (2026-06-10): https://github.blog/changelog/2026-06-09-claude-fable-5-is-generally-available-for-github-copilot/
  11. Anthropic — Claude Fable product page (2026-06-10): https://www.anthropic.com/claude/fable
  12. Simon Willison — Initial impressions of Claude Fable 5 (2026-06-10): https://simonwillison.net/2026/Jun/9/claude-fable-5/
  13. Microsoft Azure AI Foundry — Fable 5 model catalog (2026-06-10): https://ai.azure.com/catalog/models/claude-fable-5
  14. Anthropic API Docs — Migration guide Opus 4.8 → Fable 5 (2026-06-10): https://platform.claude.com/docs/en/about-claude/models/migration-guide

Verification status

Researcher-compiled 2026-06-10; flagged-UNVERIFIED items are marked inline. Independent spot-check (2026-06-10, separate agent, sources [1][2][4] fetched fresh) on five load-bearing claims: - ✅ CONFIRMED: model id claude-fable-5 exact; pricing $10/$50 (Opus $5/$25, Haiku $1/$5); effort levels incl. xhigh + max on Fable 5 and low "suitable for subagents"; Covered-Model 30-day retention / no ZDR. - ✏️ CORRECTED: "thinking disabled → 400 error" — official docs say "not supported" without specifying an error code; §2 wording adjusted.