Fable 5 and GPT 5.5: a worksheet for reviewing the Scripti code comparison

Keep the original test context attached

The 9 June 2026 description names Claude Fable 5.0, GPT 5.5 through Codex and Luke’s production SaaS, https://scripti.ai. It describes four activities: finding bugs, fixing them, redesigning the UI and a security audit. Use the chapter list to follow the bug-hunting, UI and security segments, recording proposed repairs separately from checked results.

The source also calls Fable 5 “Mythos”. Preserve that as the video’s naming claim rather than treating it as independently verified model documentation. This guide provides a viewing and review method, not access to Scripti’s private code.

Examine what a bug count contains

The description reports 23 bugs for Fable and one for Codex, including overcharging, credit race conditions and promo abuse. Those counts belong to this run. A useful comparison asks which findings are distinct, reproducible and consequential before making a judgement from the total.

1. For each reported bug, note the affected flow, expected rule and reproduction evidence.
2. Keep a proposed repair separate from a rerun demonstrating that the original failure is fixed.

Review design and security by their own criteria

Luke describes a bold pricing-page redesign from Fable versus Codex’s “live app polish”. The security examples name self-granted paid tiers, promo redemption races, a leaking debug router and rate-limit bypass. These topics require checks against actual application permissions and behaviour, not only convincing explanations.

1. Follow the pricing-page user journey after any UI change, checking clarity and preserved functionality.
2. Reproduce security findings only on systems you own or are authorised to test; record permitted and denied cases after a correction.

Retain the exception as well as the verdict

The source credits Codex on dependency hygiene and NPM advisories while expressing Luke’s preference for Claude. Keep that narrower advantage visible. Usage limits, the repository state and the chosen tasks constrain the conclusion; this account is not a complete security assessment or a timeless leaderboard.

For dependency advisories, record the affected package and suggested action separately from application-logic findings. Keep unresolved items visible even when the preferred model produces more findings overall, so the final review remains about the product’s actual problems.

The takeaway

Evaluate findings and repairs individually, and keep the video’s model preference attached to its specific codebase and date.