Breaking

Users Extract Fable 5's Instructions to Recreate Its Behavior on Opus 4.8

July 10, 2026 at 19:45 EDT

Foundation Models
Software Dev & Coding
AI Agents

A community effort to extract what users call the "operating manual" of Anthropic's Claude Fable 5 and apply it to the lower-tier Opus 4.8 is spreading ahead of the July 12 change to access limits. The goal is to preserve Fable 5's behavior on the cheaper, always-available Opus 4.8 as free access to Fable 5 narrows.

July 12 Deadline · Claude Model Access Shift

Cloning Fable 5's "Operating Manual" onto the Cheaper Opus 4.8

As free Fable 5 access narrows and credit-billing takes over after 23:59 PT on July 12, users are extracting Fable 5's system prompt and loading it into the always-available Opus 4.8 — reportedly reproducing close to 90% of its behavior.

~90%

of Fable 5 behavior said to be reproduced by loading its instruction set into Opus 4.8

1,585

lines in Fable 5's system prompt — over 120,000 characters (~27,000 tokens)

2×

Fable 5's token pricing versus Opus 4.8, input and output alike

Price per 1M tokens — Fable 5 costs twice as much

Each block = $5. Taller column = higher cost.

INPUT

Opus 4.8

$10

Fable 5

OUTPUT

$25

Opus 4.8

$50

Fable 5

SWE-Bench Pro Fable 5 ~80.3% vs Opus 4.8 ~69%

Context / Max output 1M / 128K tokens — identical

The extraction workflow

STEP 1

Prompt Fable 5 to "write the operating manual your replacement will run on"

→

STEP 2

Capture procedures, failure modes and self-tests into an instruction document

→

STEP 3

Load into Opus 4.8 via system prompt / CLAUDE.md for "Fable 5 Lite" behavior

What carries over

Reproduces ~90% of Fable 5's behavior
Makes Opus 4.8 more proactive, less low-quality output
Strong on large-codebase migrations and long-horizon agentic tasks

What prompts can't clone

"Real reasoning depth" lives in model weights, not prompts
Biology / cyber requests still fall back to Opus 4.8
Judgment gaps remain — insurance, not a full replacement

Continue reading

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