Advisor tool: Fable 5 (claude-fable-5) consistently returns error_code "unavailable"; Opus 4.8 succeeds under identical config

Status Open

Reported on v2.1.205

Maintainer reply None cached

Activity 3 comments · opened Jul 9, 2026

Description

The advisor tool fails 100% of the time (2/2 attempts, two independent sessions) when advisorModel is set to fable (claude-fable-5), while the exact same executor (Sonnet 5) succeeds immediately when advisorModel is set to opus (claude-opus-4-8).

Environment

Claude Code version: 2.1.205 (macOS)
Executor model: Sonnet 5 (claude-sonnet-5)
Advisor model under test: Fable 5 (claude-fable-5)

Steps to reproduce

Set "advisorModel": "fable" in settings (or pass --advisor fable) with executor --model sonnet.
Trigger an advisor call from within a session (interactively via the advisor tool, or headless: claude -p "<prompt instructing the model to call advisor>" --advisor fable --model sonnet).
Observe the tool result.

Expected behavior

Per the Model compatibility table, Claude Sonnet 5 → Claude Fable 5 is a documented, supported executor/advisor pair. A successful advisor_tool_result (variant advisor_redacted_result for Fable, per docs) is expected.

Actual behavior

Every attempt returns:

{"type": "advisor_tool_result_error", "error_code": "unavailable"}

Reproduced twice in independent sessions:

requestId req_011CcqsqAHekKpGEFnCpymPy, 2026-07-09T04:47:29.042Z
requestId req_011CcqtrbPQdvShwenumrYy4, 2026-07-09T05:00:37.074Z

No advisor_message iteration ever appears in usage.iterations for these calls — the advisor sub-inference appears to fail before any Fable-side tokens are consumed. This also means Fable advisor usage never shows up in account usage/billing despite advisorModel being correctly configured, which is what led me to investigate in the first place.

Control test (same executor, different advisor model)

Identical setup, only --advisor opus instead of --advisor fable:

claude -p "<explicit instruction to call advisor>" --advisor opus --model sonnet

succeeded immediately — requestId req_011CcquLsfyRnp8vEJGHTnam, 2026-07-09T05:07:00.116Z, advisorModel: "claude-opus-4-8", returned a proper advisor_result with real text content.

Why this matters

This makes the documented "Sonnet executor + Fable advisor" cost-optimization pattern (as described in Anthropic's own "the advisor strategy" post) non-functional in practice — the advisor call is attempted, correctly reaches the server, and is rejected every time, silently degrading to "no advice" while the executor continues unassisted. Since error_code: "unavailable" is documented as a generic catch-all, it isn't possible to self-diagnose further from the client side.

Ask

Could someone confirm whether Fable 5 (or Mythos 5) is currently fully enabled as an advisor model for standard accounts, or if this is a known rollout/availability gap?

View original on GitHub ↗

3 Comments

Woo7i · 1 month ago

[Bug] Filing this as a bug — not a usage question. I don't have label permissions on this repo (external contributor), so flagging here for triage: this is a 100% reproducible server-side failure (error_code: "unavailable") specific to Fable 5 as the advisor model, with a passing control test (Opus 4.8) under an otherwise identical config. Happy to provide more repro detail if needed.

aadityakbh · 1 month ago

Confirming this with a different executor, which narrows it further.

The report above pairs a Sonnet 5 executor with the Fable advisor. I hit the identical failure with Opus 4.8 (1M context) (claude-opus-4-8[1m]) as the executor and "advisorModel": "fable" in ~/.claude/settings.json:

First advisor call of the session, no prior calls, so not rate-limiting from usage.
Tool result: The advisor tool is unavailable. Do not try to use it again. — no error code, no requestId surfaced to the model, no retry hint.
Date: 2026-07-09 (same day as the repros above, so still live).

Since the failure reproduces across two different executor models (Sonnet 5 and Opus 4.8) against the same advisor model, this looks like Fable-as-advisor being unavailable outright rather than an executor/advisor pairing incompatibility. That's consistent with your observation that no advisor_message iteration ever appears in usage.iterations — the sub-inference is rejected before any Fable tokens are consumed.

Worth flagging one downstream effect for whoever picks this up: because the harness surfaces it as a terminal Do not try to use it again, an agent mid-task drops the advisor for the remainder of the session and proceeds unassisted. The degradation is silent from the user's side — they configured an advisor, are billed for none, and get no advice, with nothing in the transcript explaining why. A distinguishable error code (advisor_model_unavailable vs. the generic unavailable, per #66784) would at least make it self-diagnosable.

Workaround for anyone landing here: set "advisorModel": "opus". Advisor works immediately.

mihailmariusiondev · 1 month ago

Reproduced on Claude Code v2.1.205 (Claude Max, Linux/WSL2), with data that narrows down the cause — for us this is not Fable-specific serving, it is the deferred-tools / ToolSearch interaction reported in #73923.

Environment: Sonnet 5 main (also tested Fable 5 main), advisorModel: fable, valid pairing, no gateway.

Observations:

Interactive sessions: first advisor call → Advisor unavailable (unavailable). Once failed, the session stays latched off (matches #75903 / #74530) — /advisor re-selection does not recover it, and --continue inherits the broken state (#73128).
Headless, same account/settings/repo, minutes apart: claude -p --model sonnet asking for one advisor call succeeded 5/5 (also with --model claude-fable-5).
A/B with the tool-search env var, both headless, same prompt shape:

ENABLE_TOOL_SEARCH=true + load a deferred tool via ToolSearch, then call advisor → The advisor tool is unavailable. Do not try to use it again.
ENABLE_TOOL_SEARCH=false (all tools upfront) → advisor succeeds.

Workaround verified: setting "env": {"ENABLE_TOOL_SEARCH": "false"} in settings.json fixes Fable-as-advisor in fresh interactive sessions (confirmed working). Previously-failed sessions remain latched and must be abandoned.

So at least for this account, "Fable advisor unavailable" = deferred tools present in the request (default in interactive sessions) breaking the advisor server tool, plus the per-session latch. Opus-as-advisor appearing to "work" may just have different tolerance on that path.