OpenAI Uses GPT-5.6 Soul to Automate Model Efficiency Gains

Recursive Efficiency Gains

OpenAI utilized their frontier model, GPT-5.6 Soul, to perform automated research into their own infrastructure, resulting in significant cost and performance improvements. By analyzing production traffic and testing architectural changes, the model identified ways to optimize the forward pass, rewrite production kernels, and improve speculative decoding. This process mirrors the concept of recursive self-improvement, where a high-capability model iterates on its own efficiency without human intervention.

Pricing and Performance Shifts

Following these optimizations, OpenAI dropped the price of GPT-5.6 Luna to $0.20 per million input tokens and $1.20 per million output tokens. While GPT-5.6 Soul did not receive a direct price cut, its fast mode now offers 2.5x speed for 2x the cost, effectively lowering the cost per task. Comparative data from Artificial Analysis indicates that GPT-5.6 Luna now achieves a cost per task of approximately $0.06, significantly undercutting competitors like GLM-5.2 Max ($0.26) and Claude Opus 5 ($0.40) while maintaining comparable intelligence scores.

Strategic Implications

This development suggests a shift in the business model for frontier labs: training massive, expensive models to serve as internal "researchers" that generate highly efficient, cost-effective versions for public consumption. By automating the discovery of kernel improvements and routing heuristics, OpenAI creates a compounding advantage that may be difficult for open-source or smaller competitors to match. The strategy highlights the growing importance of cost-per-task metrics over simple per-token pricing, as models like Kimmy K3 demonstrate that lower token costs do not always equate to lower total task costs due to varying token efficiency.