AI
OpenAI's new GPT-6 models cut token costs in half, but smarter caching may be the real savings lever
Prompt caching improvements promise to reduce API costs even further by letting developers reuse processed context without restarting computations.
The New Stack
OpenAI's latest release of GPT-6 Sol and Luna brings two simultaneous cost reduction strategies: slashing token prices by 50% compared to their GPT-5.6 predecessors, and overhauling how the models handle cached prompts to minimize redundant processing. Together, these changes aim to make agent-based AI applications substantially cheaper to run at scale.
Bigger discounts on fresh tokens, plus intelligent reuse#
The headline pricing drops are significant—Sol input tokens fell from $4 to $2 per million, while Luna dropped to just $0.10 from $0.20. Output token reductions track similarly steep. But OpenAI argues the real economic opportunity lies beneath these numbers: by making cached context stick across more task variations, agents can avoid reprocessing the same information repeatedly.
The company reports that cached input reads carry a 90% discount, matching earlier models. What has changed is the frequency of cache hits. "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 company stated.
Caching without friction#
Previously, adjusting reasoning effort or changing which tools an agent could access would force the cache to reset, losing all prior savings. GPT-6 Sol and Luna now preserve cached context even when these settings shift. This flexibility lets developers dial up computational effort for tricky reasoning steps while keeping earlier work cached, then dial it back for simpler decisions—all without throwing away what the model has already processed.
The result cuts both latency and token consumption. Developers no longer face a false choice between flexible task handling and cache efficiency.
New transparency tools for optimization#
A fresh Prompt Caching Dashboard lets developers inspect caching behavior in detail—showing how much input gets reused over time and flagging opportunities where caching could improve. Rather than treating cache performance as a black box, the tools aim to make it measurable and optimizable. According to OpenAI, GitHub has used these improvements to cut fresh token processing by over 50% across billions of recent requests.
