Breaking

Anthropic Details Pattern Using Fable 5 as Advisor and Sonnet 5 as Executor

July 7, 2026 at 17:27 EDT

AI Agents
Foundation Models
Software Dev & Coding

Anthropic's official developer team on July 7, 2026 published a hybrid usage pattern that employs its top-performing Claude Fable 5 model as an "advisor" while leaving execution to the cheaper Sonnet 5, aiming to improve cost efficiency by billing most tokens at the lower executor rate.

June 2026 · Anthropic

The Advisor Pattern: Premium Insight, Budget Bill

Anthropic pairs its top-tier Fable 5 as an on-demand "advisor" with the cheaper Sonnet 5 executor handling most tokens — keeping high-end performance while cutting operating cost.

96%

of performance kept
(BrowseComp)

46%

of the price
(BrowseComp)

token context
(Fable 5)

API Price per Million Tokens

Executor Sonnet 5 processes the bulk of tokens at ~1/5 the cost

$10 / $50

Fable 5 — advisor

input / output

$2 / $10

Sonnet 5 — executor

input / output

SWE-bench Pro

Sonnet 5 + Fable 5 advisor holds ~92% of the score at ~63% of the price.

BrowseComp

Fable 5 orchestrator + Sonnet 5 worker keeps 96% performance at 46% of the price.

How the loop runs

Sonnet 5

executor advances the task

→

Fable 5

called only when guidance needed

→

Result

most tokens billed at the cheaper rate

What works

Large codebase migrations cut from "months to days"
Cheap access to Fable 5's insight via the advisor pattern
Sonnet 5 executor + Fable 5 advisor rated cost-efficient

Open concerns

Fable 5 is expensive and token-hungry — use cheaper sub-agents
Safety classifiers fire on benign queries, falling back to Opus 4.8
Export-control access limits and advisor implementation bugs

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.