What happened
Charles Azam benchmarked Anthropic's Claude Fable 5 against OpenAI's GPT-5.6 Sol on an NP-hard fiber-network design problem, the KIRO challenge. Fable 5 consistently delivered superior solutions, achieving a mean score of 32,386 compared to Sol's 34,261, with its best solution at 31,934. The models' native /goal mode, designed to guide optimisation, frequently degraded mean performance for both models, despite winning four of six individual trials, indicating it is not a universal performance enhancer.
Why it matters
For architects and procurement teams evaluating frontier models for complex optimisation, Fable 5's consistent performance on an NP-hard problem demonstrates that core model intelligence and stability are critical. The /goal mode, while winning individual trials, often worsened mean results, indicating that specific control mechanisms can introduce performance regressions rather than guarantee improvement. This contrasts with recent reports of xAI Grok 4.5 leading coding performance, highlighting varied strengths across leading models. Teams should prioritise empirical validation over feature claims for mission-critical applications.
