Fable Brain

Model Selection for Agent Workloads

5 min read

Model Selection for Agent Workloads

For agent work in July 2026, default to Sonnet 5 at 15, escalate to Fable 5 at 50 only when the task needs frontier capability and can tolerate refusal-fallback plumbing, and keep Haiku 4.5 for high-volume subagent steps.

What it is

Contradiction

The claude.ai export’s lineup (System Prompt Export 2026-07, L21-23) lists claude-sonnet-4-6 and omits Mythos 5, while the current docs list claude-sonnet-5 and claude-mythos-5. The export is a 2026-06-09 snapshot; select models from the dated docs page, not the corpus.

How it works

  • Capability axis: Fable 5 is Mythos-class, above the Opus class; Anthropic states its capabilities exceed any previously generally available model (https://www.anthropic.com/news/claude-fable-5-mythos-5, published 2026-06-09).
  • Cost axis is not just the sticker price: Fable 5 uses the Opus 4.7 tokenizer, producing roughly 30% more tokens than pre-4.7 models for the same text, which inflates both input and output bills.
  • Caching axis: Fable 5 offers a 90% prompt-caching input discount, which rewards agent loops with large stable system prompts; US-only inference costs 1.1x.
  • Thinking axis: adaptive thinking is always on for Fable 5 and cannot be disabled; depth is steered with the effort parameter, and raw chain of thought is never returned.
  • Refusal axis: only Fable 5 carries dual-use classifiers (cyber, bio, frontier_llm, reasoning_extraction). Refusals return HTTP 200 with stop_reason refusal and can fall back to Opus 4.8 server-side, via SDK middleware, or by manual retry with fallback credit.
  • Compliance axis: Fable 5 is a Covered Model with mandatory 30-day retention and no zero-data-retention option; workloads contractually requiring ZDR must use a non-covered model.
  • Context axis: Fable 5, Opus 4.8, and Sonnet 5 all offer 1M-token windows; Haiku 4.5 stops at 200k, which bounds its role to scoped subtasks.
  • Field report: Simon Willison’s day-one testing found Fable 5 slow and expensive but exceptionally capable, spending $110.42 in about 5.5 hours; “The challenge is finding tasks that it can’t do.” (https://simonwillison.net/2026/Jun/9/claude-fable-5, published 2026-06-09).

Best practice

  • Start agent pipelines on Sonnet 5 and promote individual stages to Fable 5 only when evaluation shows a capability gap; the price gap is 3x or more per token. PRACTITIONER
  • Route high-volume, low-stakes subagent calls (classification, extraction, triage) to Haiku 4.5 and keep its 200k window in mind when packing context. PRACTITIONER
  • If a stage runs Fable 5 unattended, wire the server-side fallbacks parameter or SDK refusal middleware to Opus 4.8 before shipping. EVIDENCE-BASED
  • Recompute budgets with the Opus 4.7 tokenizer’s roughly 30% token inflation before comparing Fable 5 quotes against 4.5-era baselines. EVIDENCE-BASED
  • Exploit the 90% prompt-caching discount by freezing system prompts and tool schemas across loop iterations. EVIDENCE-BASED
  • Exclude Fable 5 from any workload that requires zero data retention; the Covered Model designation makes 30-day retention mandatory. EVIDENCE-BASED
  • Treat day-one speed and cost impressions as provisional; they come from one practitioner on launch day. CONTESTED

Pitfalls

  • Choosing Fable 5 by default “because it is the best”: latency, token inflation, and refusal plumbing make it the wrong default for most agent steps.
  • Ignoring refusal handling because triggers average under 5% of sessions; security and bio-adjacent agent workloads concentrate well above the average.
  • Selecting Opus 4.8 as the fallback target but forgetting rerouted requests are billed at Opus prices, not Fable prices.
  • Copying the model lineup out of the system prompt export; it is a stale launch-day snapshot.
  • Assuming availability is constant: Fable 5 was suspended 2026-06-12 under export controls and restored 2026-07-01, so single-model architectures carry regulatory risk.

Sources

Next actions

  • Build a small eval set from this vault’s real agent tasks and measure Sonnet 5 vs Fable 5 stage by stage.
  • Re-check Sonnet 5 introductory pricing expiry on the models overview page in August 2026.
  • Record observed refusal rates per workload to replace the launch-day sub-5% average with local data.

Graph View

Backlinks