Back to Blog List

Claude Fable 5 Guide: Benchmarks, Cost, and Fallback Behavior

June 9, 2026
Updated August 12, 2026
Updated August 2026 developer guide to Claude Fable 5: $10/$50 pricing, anthropic_default_fable_model setup, refusal fallback FAQ, and when to route vs Opus.
Claude Fable 5 Guide: Benchmarks, Cost, and Fallback Behavior
claude fable 5
anthropic
anthropic_default_fable_model
claude code
pricing
benchmarks

Last updated: August 12, 2026. Claude Fable 5 is Anthropic's generally available Mythos-class model for long-horizon coding and knowledge work. This guide refreshes pricing, Claude Code environment variables (including ANTHROPIC_DEFAULT_FABLE_MODEL / anthropic_default_fable_model), fallback behavior, and practical routing advice. Pricing and product policy can change—confirm on Anthropic's official pages before you ship.

Quick facts

| Item | Value (as of sources checked Aug 2026) | | --- | --- | | API model id | claude-fable-5 | | Class | Mythos-class (general access; Mythos 5 is the restricted sibling) | | API pricing | $10 / M input tokens, $50 / M output tokens | | Prompt caching | Anthropic documents a large cached-input discount (commonly cited ~90% off input for cache hits)—verify live docs | | Context | High long-context surface (OpenRouter listings show 1M-class context; confirm in current Anthropic model table) | | Data retention | Mythos-class traffic typically requires 30-day retention for safety monitoring (not training)—confirm policy text |

What developers should know before shipping Fable 5

Fable 5 is not a drop-in "always better, same cost" upgrade from Opus-class defaults. It is a premium, long-horizon model:

  • Stronger on hard multi-step coding / analysis workloads (Anthropic's launch narrative and early customer stories).
  • 2× the common Opus 4.8-style sticker rates many teams used as a baseline ($10/$50 vs roughly $5/$25 class pricing for Opus 4.8 in public comparisons).
  • Safety classifiers can refuse and hand work to a fallback model (often Opus 4.8)—your integration must handle this on the API.

Pricing math (verify live)

Public Anthropic launch materials price Fable 5 and Mythos 5 at:

$10 per million input tokens
$50 per million output tokens

Worked example for one agentic task with 200K input and 50K output:

Uncached ≈ (0.2 × $10) + (0.05 × $50) = $4.50

With aggressive prompt caching on a stable prefix, billable input can drop substantially on later turns—design prompts so the cacheable prefix is stable and volatile state sits at the end.

Other footnotes seen in public docs / write-ups:

  • US-only inference may carry a 1.1× multiplier where data residency is required.
  • Consumer/subscription surfaces may count Fable usage with higher rate multipliers than Opus—check plan FAQs.
  • Batch / cache prices: confirm on the official pricing page (do not hardcode forever).

Setting ANTHROPIC_DEFAULT_FABLE_MODEL (Claude Code)

Many developer searches look like anthropic_default_fable_model (lowercase) or the env var form ANTHROPIC_DEFAULT_FABLE_MODEL. Claude Code uses this to pin which model id counts as the Fable tier default (and related selector UI on first-party / gateway providers).

Shell (macOS / Linux)

# Pin Fable tier to official Fable 5
export ANTHROPIC_DEFAULT_FABLE_MODEL="claude-fable-5"

# Optional: pin Opus fallback tier used when Fable refuses / you configure dual defaults
export ANTHROPIC_DEFAULT_OPUS_MODEL="claude-opus-4-8"   # use the exact id from your Anthropic account docs

# Optional display metadata supported by some Claude Code builds
export ANTHROPIC_DEFAULT_FABLE_MODEL_NAME="Claude Fable 5"
export ANTHROPIC_DEFAULT_FABLE_MODEL_DESCRIPTION="Mythos-class long-horizon coding"

claude

Make it permanent

# ~/.zshrc or ~/.bashrc
export ANTHROPIC_DEFAULT_FABLE_MODEL="claude-fable-5"
export ANTHROPIC_DEFAULT_OPUS_MODEL="claude-opus-4-8"

Project / gateway JSON (pattern)

Some gateway or multi-model setups store the same keys in config:

{
  "env": {
    "ANTHROPIC_DEFAULT_FABLE_MODEL": "claude-fable-5",
    "ANTHROPIC_DEFAULT_OPUS_MODEL": "claude-opus-4-8",
    "ANTHROPIC_DEFAULT_SONNET_MODEL": "claude-sonnet-4-6",
    "ANTHROPIC_DEFAULT_HAIKU_MODEL": "claude-haiku-4-5"
  }
}

Use the exact model ids your provider exposes. If you route through a third-party gateway, the Fable default might look like anthropic/claude-fable-5 or a vendor alias.

Windows PowerShell

$env:ANTHROPIC_DEFAULT_FABLE_MODEL = "claude-fable-5"
$env:ANTHROPIC_DEFAULT_OPUS_MODEL = "claude-opus-4-8"
claude

Fallback, refusals, and troubleshooting FAQ

Why did Fable 5 return a refusal or an Opus answer?

Fable 5 runs safety classifiers. On Claude apps, flagged traffic may auto-reroute (often to Opus). On the API, you commonly receive an HTTP 200 with stop_reason: "refusal" (plus classifier metadata in current API shapes)—your code must handle it. Anthropic documents fallback parameters so you can retry on Opus without paying Fable rates for the rerouted work (confirm current billing language).

Claude Code does not switch models when Fable fails

If either ANTHROPIC_DEFAULT_FABLE_MODEL or ANTHROPIC_DEFAULT_OPUS_MODEL cannot be resolved to a real model id on the active provider, Claude Code may not switch and the request ends in a refusal you clear with /model or by fixing env vars.

Checklist:

  1. Print the env vars in the same shell you launch Claude Code from.
  2. Confirm the model id exists for your API key / workspace.
  3. Avoid stale aliases after Anthropic renames snapshots.
  4. For third-party base URLs, set ANTHROPIC_BASE_URL / auth tokens and matching model ids the gateway understands.

anthropic_default_fable_model is unset or ignored

# Debug: is the variable visible to the process?
echo "$ANTHROPIC_DEFAULT_FABLE_MODEL"

# Common fixes
unset ANTHROPIC_MODEL   # a global pin can override tier defaults in some setups
export ANTHROPIC_DEFAULT_FABLE_MODEL="claude-fable-5"
hash -r
claude

If you use a process manager, IDE task, or Docker container, export the variable inside that environment, not only in an unrelated terminal tab.

High burn on Max / Team plans

Fable is premium and often token-heavy on agentic loops. Mitigations:

  • Route only hard tasks to Fable; keep defaults on Opus/Sonnet.
  • Stabilize system prompts for cache hits.
  • Cap max turns / max tokens per agent run.
  • Log stop_reason and token usage per task.

Benchmarks and capability framing

Anthropic's launch materials position Fable 5 ahead of prior Opus-class defaults on hard long-horizon coding and analysis suites (including SWE-Bench Pro style evaluations in launch posts). Exact point gaps move as baselines update.

For production decisions:

  1. Pick 3–5 real tasks from your backlog (migration, multi-file bug, research synthesis).
  2. Measure task success, human edits, wall time, and $ per success—not only official charts.
  3. Keep an Opus (or cheaper) path for routine work.

If a third-party site quotes a score we cannot confirm on Anthropic's primary docs, treat it as unverified.

Claude Fable 5 vs Opus: routing rule of thumb

| Use Fable 5 when | Use Opus / cheaper when | | --- | --- | | Multi-hour agent runs, large migrations | Short, well-scoped edits | | Cost of failure ≫ token cost | High volume chatty loops | | Deep analysis / high-fidelity coding | Latency-first UX |

Default architecture many teams converge on: Fable as orchestrator / hard-reasoner, smaller models for mechanical tool steps.

30-day retention (compliance)

Mythos-class models including Fable 5 are associated with 30-day safety monitoring retention on first- and third-party surfaces (not for training, with deletion after the window in typical policy language). Update DPAs / privacy copy before enterprise rollout. Some enterprise products keep Fable off by default until the retention term is acknowledged.

Decision checklist (updated)

  1. Confirm $10 / $50 (or current) pricing in your region and channel.
  2. Set ANTHROPIC_DEFAULT_FABLE_MODEL (and Opus default) correctly for Claude Code.
  3. Implement API refusal / fallback handling before production.
  4. Score $ per completed hard task during a pilot—not vibes.
  5. Document retention for legal/security review.

Frequently asked questions

What is Claude Fable 5?

Anthropic's generally available Mythos-class model for long-horizon coding and knowledge work, API id typically claude-fable-5, launched June 9, 2026 alongside Claude Mythos 5.

How much does Claude Fable 5 cost?

Public list pricing at launch: $10 per million input tokens and $50 per million output tokens. Always re-check Anthropic's pricing page.

What is anthropic_default_fable_model?

It is the Claude Code / environment configuration key (env: ANTHROPIC_DEFAULT_FABLE_MODEL) that selects which model id is treated as the Fable tier default. Set it to claude-fable-5 (or your gateway's equivalent) so Fable selection, fallbacks, and UI labels resolve correctly.

Is Fable 5 better than Opus?

On hard, long-horizon tasks Anthropic positions Fable ahead; on routine work Opus (or cheaper) is often the better value. Route by task, not by brand.

Does Fable 5 require special data retention?

Plan for Mythos-class 30-day safety retention semantics and update compliance docs accordingly.

Sources for this refresh: Anthropic Claude Fable 5 / Mythos 5 launch posts and model docs (pricing $10/$50), Claude Code environment variable references for ANTHROPIC_DEFAULT_FABLE_MODEL, and public developer write-ups on fallback/refusal behavior. Re-verify against Anthropic before production commits.

Back to Blog List