We benchmarked four production fine-tuned models against Claude Opus 5.5, Claude Fable 5.1, GPT-6 Astra and GPT-6.1 Sol on held-out customer tasks. Accuracy, cost, latency and fallback rate.
