OpenAI's cheapest model Luna slashes costs; GPT-5.6 Sol triples ARC-AGI v3 score with API config fix

OpenAI's Efficiency Push: Luna Cuts Costs, Sol Fixes ARC AGI Through API Configuration

OpenAI is pursuing a two-pronged strategy to expand model access: making cheaper inference available at scale while simultaneously fixing performance gaps that turned out to be harness problems, not fundamental model limitations.

Luna, OpenAI's cheapest model, sits on a new part of the efficiency frontier—cheaper than many open-source alternatives when measured by cost per task rather than raw token price. The distinction matters. A model that costs half as much per token but requires 10 times the token count winds up costing more overall. Luna avoids that trap by being genuinely efficient.

The ARC AGI comeback is sharper. GPT-5.6 Sol initially underperformed on ARC AGI v3, scoring poorly enough to raise questions about whether the model had hit a wall on reasoning tasks. OpenAI investigated and found the problem wasn't the model—it was the API harness. Enabling two specific API settings tripled the score while reducing output tokens by six times.

This pattern has become familiar over the past 18 months: the model's raw capability often exceeds what the integration reveals. A badly configured harness can constrain performance dramatically. The implication is uncomfortable for anyone betting on a single monolithic model doing everything—the real work is in how you ask it to think, what you let it remember across tasks, and how you structure the problem.

The broader pattern OpenAI is signaling is that frontier improvements now come from both model capability and integration sophistication. Cheaper models with proper configuration can compete on cost per task. Capable models with the right API settings unlock performance that looked missing.