OpenAI cuts GPT-5.6 Luna and Terra prices as it broadens the AI cost-performance race

OpenAI has lowered prices for two GPT-5.6 models while introducing a faster processing option for its highest-end API model, sharpening competition over the cost of deploying capable AI systems. The company said GPT-5.6 Luna, positioned for high-volume workloads, is now 80% cheaper, while the balanced Terra model is 20% cheaper. GPT-5.6 Sol remains at its existing price but gains a Fast mode that OpenAI says can deliver up to 2.5 times the speed of standard processing at a premium.

The change is consequential because enterprise AI economics are moving beyond headline benchmark results. Teams increasingly assemble workflows that use different models for planning, execution, review and tool use. A reduction in the cost of a smaller, capable model can determine whether an agent is practical for continuous document processing, support operations or background coding work rather than only occasional demonstrations.

OpenAI says the lower prices reflect improvements across models, inference systems and the software that routes work through them. It also points to gains from more efficient context handling, an important part of the calculation for agents that must retain state across multiple steps. The company frames the update as an effort to let customers choose the amount of intelligence, latency and reliability appropriate to each part of a job.

For buyers, the immediate comparison will be against the total cost of a completed outcome rather than a single token price. A cheaper model that needs repeated corrections can erase a price advantage, while a faster frontier model may be justified for a time-sensitive decision. The new lineup gives builders another reason to separate routine execution from more difficult reasoning rather than sending every request to one model.

The announcement also puts pressure on competing platforms to explain their own trade-offs among capability, speed and price. OpenAI's figures are company statements and their practical effect will depend on real production quality, quota rules and regional availability. The next test is whether lower serving costs translate into broader, durable adoption of multi-step AI workflows.