Opus 4.8

The Shift from Code to Process Evaluating frontier AI models on simple code generation is no longer useful. Every major model can generate a working code block and pass basic tests. To find the real dividing line, we must throw these systems into legacy codebases with real-world problems. Testing GPT-5.6 Soul, GPT-5.5, Fable 5, and Opus 4.8 on a real, open-source bug from the BookStack repository reveals that the ultimate differentiator is no longer the final syntax—it is the engineering methodology. Soul Masters Test-Driven Development When tasked with fixing a bug in the old BookStack application, GPT-5.6 Soul stood entirely alone in its execution flow. Before editing a single line of production code, it generated a failing integration test to reproduce the reported issue. This classical Test-Driven Development (TDD) approach provides a safety net that other models ignored. Rivals like GPT-5.5 and Opus 4.8 Medium bypassed this step, diving straight into modifying the export formatter without verifying the failure state first. Fable 5 Prioritizes Security and Permissions While GPT-5.6 Soul focused on TDD, Fable 5 expanded the scope toward project security. Operating at a medium reasoning level, it did not just fix the core bug. It identified potential authorization vulnerabilities within the export logic, automatically introducing permission-checking tests. This ability to think wider than the prompt makes Fable 5 a strong candidate for architecture and planning, even if it lacks the TDD-first instinct of Soul. Opus 4.8 Rewrites for Future Scalability Opus 4.8 High Effort approached the task with maximum thoroughness. It ran for the longest duration, refactored a regex pattern to make it reusable outside the DOM context, and triggered the entire suite of 1,500 tests. By widening the scope, it guarded against regression across the whole codebase, though at a significant cost. Cost Versus Capability GPT-5.6 Soul delivers the best overall balance of price and logic. It matches GPT-5.5 in token pricing while offering vastly superior, TDD-driven execution. Opus 4.8 High Effort and Fable 5 both ran up API bills over twice as expensive as Soul, without delivering twice the utility. For developers seeking automated pull requests they can actually trust, Soul's test-first methodology makes it the most reliable driver in the fleet.