OpenAI released GPT-6 Sol and Luna this week with 50% lower token prices than their predecessors, but the company's real cost-cutting innovation may be hiding in a less obvious place: improved prompt caching that lets AI agents reuse previously processed information instead of reprocessing it from scratch. This dual approach to affordability signals a shift in how AI companies think about making their models economical for developers building complex, long-running tasks.

What Exactly Is Prompt Caching and Why Should Developers Care?

Prompt caching is a technique that stores context the model has already processed, so it doesn't have to do that work again on the next request. Think of it like a chef prepping ingredients once instead of chopping vegetables for every single dish. When an AI agent works on a multi-step task, it often needs to reference the same information repeatedly. Without caching, the model processes that information fresh every time, burning through tokens and adding latency. With caching, the model retrieves the stored context instantly and at a steep discount.

OpenAI's new GPT-6 models come with three major caching improvements designed to make this process more efficient and transparent. The company says these upgrades have already reduced the share of prompt tokens requiring fresh processing by more than 50% across billions of requests to OpenAI models over the past several months.

How to Optimize Your AI Costs With Caching and Pricing Improvements

• Leverage Higher Default Cache-Hit Rates: GPT-6 Sol and Luna automatically preserve more previously processed context, meaning agents reuse more information without developers having to manually configure caching. This reduces both latency and token costs, with cached input tokens discounted by 90% compared to fresh processing.
• Adjust Reasoning Effort Without Breaking the Cache: The new models let developers scale reasoning effort up or down based on task difficulty and change tool availability without disrupting earlier cached context. This flexibility means you can optimize performance for each step without losing the cost benefits of cached information.
• Monitor Caching Performance With the New Dashboard: OpenAI introduced a Prompt Caching Dashboard that shows developers exactly how much context is cached, how that amount changes over time, and flags missed caching opportunities. This visibility helps teams actively measure and optimize cache reuse rather than leaving performance hidden behind the scenes.

How Much Cheaper Are the New Models, Really?

The headline numbers are striking. GPT-6 Sol input tokens now cost $2 per million, down from $4 per million for GPT-5.6 Sol. Output tokens dropped to $10 per million from $20. Luna saw even steeper cuts: input tokens fell from $0.20 to $0.10 per million, and output tokens dropped from $1.20 to $0.50 per million. But these price cuts alone don't tell the full story.

The real leverage comes from combining lower token prices with improved caching. When an agent reuses cached context, it pays only 10% of the normal input token price. GitHub, which uses OpenAI's models extensively, reported that these caching improvements have reduced the share of prompt tokens requiring fresh processing by more than 50% across billions of requests. For developers running agents on long, complex tasks, that compounds into substantial savings.

Why Caching Matters More Than You Might Think

As AI agents tackle increasingly longer and more complex tasks, the economics of token pricing alone become less relevant. An agent working on a research project might need to reference the same document, conversation history, or knowledge base repeatedly across dozens of steps. Without efficient caching, each step reprocesses that context. With it, the agent retrieves stored context instantly and cheaply.

OpenAI's approach signals that the company understands a fundamental shift in AI economics. Cheaper models alone won't solve the cost problem for developers building production agents. The real savings come from reducing how often the same context needs to be reprocessed. By improving cache-hit rates by default, offering flexibility to adjust reasoning and tools without breaking the cache, and giving developers visibility into caching performance, OpenAI is tackling agent costs from both angles: charging less for fresh processing and reducing how often fresh processing is needed at all.

For developers evaluating AI platforms, this suggests a new lens for comparison. Don't just look at token prices. Ask how well the platform caches context, how transparent its caching diagnostics are, and whether you can adjust model behavior without losing cached work. Those details may determine whether your AI agent costs $100 or $1,000 per month.