Description: A viral setup lets Fable 5.2 think, Sonnet 5.5 code, Opus 5.5 review, and Haiku 5.5 research. Here's what the official advisor and pricing data really show.
Check the invoice before assembling an AI executive board
Fable thinks. Sonnet builds. Opus audits. An army of Haikus investigates. The popular Claude Code setup sounds like a company in a terminal: a genius strategist, a tireless engineer, a stern reviewer, and a research department that never takes lunch.
Division of labor makes sense. But there is no evidence that this precise four-model arrangement is the cheapest and most effective for every task. Turning on the official experimental Advisor does not automatically deploy an independent Opus reviewer or a Haiku research swarm.[1]
Four job titles are not four automatic workers
Fable is suited to hard architectural decisions, recurring failures, and critical final checks. Sonnet can implement routine code changes. Opus can independently review important code. Haiku fits high-volume classification and extraction.
But advisors, subagents, and independent reviews are different mechanisms. The main model chooses when it needs advice. Additional workers require their own tasks and instructions. Listing four employees on an organizational chart does not mean all four have clocked in.[1]
“80 times cheaper” refers to a unit price
The standard API prices checked on November 14, 2026, per million input tokens were $8 for Fable, $4 for Opus, $2 for Sonnet, and $0.10 for Haiku when a prompt contains at most 100,000 input tokens. Output rates were $40, $20, $10, and $0.50 respectively.[3]
So the token rates differ by 80:1 between Fable and Haiku in that band. For Haiku requests above 100,000 input tokens, the listed base rates rise to $0.50 for input and $2.50 for output. Actual job costs also include repeated context, caching, consultations, failures, and retries.
One-eightieth the token price does not mean one-eightieth the cost per finished feature. Pencils are cheaper than structural engineers. Two hundred pencils will not design a bridge.
What /advisor actually does
A supported Claude Code installation can use:
/model sonnet
/advisor fable
Advisor is experimental. Availability depends on the platform, model entitlement, and sometimes consent to spend extra Fable usage credits. Compatible desktop sessions support it. Enabling it does not prove that any consultation occurred; the main model decides when to ask.[1]
It can make sense during major refactoring, repeated debugging failures, or important decisions before completion. Asking the chairman to authorize a filename change is simply corporate bureaucracy.
“3.4× savings” and “198 tasks with no advice” are not universal findings
The viral 3.4× cheaper claim has not been verified as a published result for this exact four-model setup. In a different Sonnet-plus-Opus Advisor evaluation, Anthropic reported an 11.9% cost reduction per task and a 2.7-percentage-point improvement. A separate coding comparison found an Advisor setup cost about 2.1 times the solo baseline with only a small gain.[2][5]
Official material also describes a 20-problem Sonnet 5.5 check with zero consultations and a distinct 300-task chart-reading test with one consultation. Neither establishes a universal “198 tasks” rule. Call rates depend on task, model, instructions, and effort setting.[2]
The first savings are less glamorous
Anthropic's cost guide starts with prompt caching, dropping unnecessary inputs, and tuning the reasoning effort of one model. Retrying only failed jobs with a stronger setting can also help. Multi-model orchestration must beat that baseline rather than merely look impressive.[2]
Breaking huge projects into smaller jobs may pay off. For short jobs, coordination and checking can cost more than the work itself. Congratulations: payroll is down, but meetings are now the company’s main product.
Choose the winner with a real test
Pick 10–20 representative development tasks. Compare A: Opus alone, B: Sonnet plus Fable Advisor, C: Fable plans then Sonnet implements. Set success criteria first: tests pass, extracted record counts match, or the bug does not return.
Track total cost, successful completions, elapsed time, retries, and actual Advisor calls. In a data-heavy bot, use deterministic code for arithmetic, small models for fuzzy categorization, and powerful models for genuinely difficult decisions. Independently validate outcomes before real-money decisions.
The final board meeting
“Fable is CEO, Sonnet heads engineering, Opus audits, and 100 Haikus do research.” Impressive. What did they deliver?
“A million tokens of meeting minutes.”
Count correctly completed jobs per dollar, not fancy job titles per prompt.
