OpenAI's GPT-6 Models Cut Costs in Half, But Smarter Caching May Be the Real Savings

On Tuesday, OpenAI introduced GPT-6 Sol and Luna, positioned as lower-cost alternatives to the flagship GPT-6 Astra. While the headline story centers on API pricing that dropped 50% relative to their GPT-5.6 predecessors, the company is equally focused on a quieter efficiency gain: smarter caching that preserves reusable context even when agents adjust reasoning intensity or swap available tools.

The new models represent a middle ground between affordability and capability. Sol and Luna perform closer to Astra's alignment standards than the earlier GPT-5.6 Sol managed, though they remain below the top-tier model's performance. OpenAI frames the dual approach to cost reduction as essential: cheaper per-token rates combined with architectural improvements that reduce how often the same context must be processed anew.

Caching that survives model adjustments

Prompt caching itself is not new to these models, but GPT-6 Sol and Luna introduce refinements aimed at keeping previously processed context reusable as tasks evolve. OpenAI stated: "We've improved prompt caching for GPT‑6 to deliver higher cache hit rates by default, helping agents reuse more context, respond faster, and benefit from discounts of 90% on cached input-token reads."

The 90% discount on cached input tokens matches existing GPT-5.6 pricing, but the meaningful shift lies in how frequently the cache remains valid. By reusing work already performed, the model avoids reprocessing identical context for each request, cutting both latency and token consumption.

A key innovation is that developers can now adjust reasoning effort and tool availability without invalidating the cache. An agent can increase reasoning intensity for difficult steps or change which tools are available based on task requirements while preserving earlier cached context. This flexibility serves both speed and cost.

Visibility into caching performance

GPT-6 Sol and Luna ship with a Prompt Caching Dashboard, giving developers insight into how much context gets reused and how that metric changes over time. The tool identifies missed caching opportunities, helping teams understand where efficiency gains are possible.

Rather than leaving cache performance opaque, OpenAI is making it measurable and actionable. According to the company, GitHub has reported that these improvements have cut the proportion of prompt tokens requiring fresh processing by more than 50% across billions of requests over the past several months.

Pricing breakdown

The token price reductions are substantial. GPT-6 Sol now costs $2 per million input tokens and $10 per million output tokens, down from GPT-5.6 Sol's $4 and $20 respectively (GPT-5.6 Sol pricing is promotional). GPT-6 Luna dropped to $0.10 per million input tokens and $0.50 per million output tokens, compared to GPT-5.6 Luna's $0.20 and $1.20.

Cost reduction from multiple angles

As AI model providers intensify price competition, the industry is learning that cheaper models alone cannot solve cost problems. OpenAI's strategy with Sol and Luna demonstrates a two-pronged approach: reducing the price of fresh token processing while simultaneously minimizing how often context must be reprocessed.

As agents tackle longer and more complex tasks, pressure to optimize both dimensions will likely grow. The combination of lower token rates and improved caching efficiency represents OpenAI's answer to the mounting pressure on AI application economics.