OpenAI slashes GPT-5.6 prices weeks after launch, citing infrastructure gains and Chinese competition

OpenAI has cut API prices for two of its GPT-5.6 models just three weeks after their introduction, signaling a shift in how the industry prices newer capabilities. The Luna variant dropped 80% in cost and Terra fell 20%, while the flagship Sol model remained unchanged. CEO Sam Altman called the move part of a push to "offer the best price/intelligence tradeoff at every level."

Luna now costs $0.20 per million input tokens (down from $1) and $1.20 per million output tokens (from $6). Terra is $2 and $12 per million input tokens and output tokens respectively, compared to $2.50 and $15 previously. Sol pricing stayed flat at $5 and $30.

The speed of these cuts is striking. Most AI vendors hold prices steady for months after launching a new model family, but OpenAI moved in less than a month, suggesting underlying pressure driving the decision.

Infrastructure efficiency enables lower costs#

OpenAI attributed the price drops to improvements across its inference stack. Engineers rewrote GPU kernels to reduce serving costs by roughly 20%, redesigned Sol's speculative decoding to make token generation about 15% faster, and optimized prompt caching to cut redundant computation in multi-step agent workflows. Each efficiency gain translated directly into lower customer costs.

These technical advances had been signaled a day before the pricing announcement when OpenAI published an engineering overview of its optimization work.

Agentic AI and token economics reshape priorities#

Developers running agentic systems are increasingly focused on inference costs because these applications often trigger dozens or hundreds of model calls to complete a single task. In such scenarios, serving efficiency matters more than small benchmark improvements between competing models. This has shifted vendor focus from pure performance to cost-per-capability.

Chinese competitors drive global pricing pressure#

The underlying driver appears to be intensifying competition from overseas, particularly lower-cost open-weight models from Chinese AI companies like Moonshot. These alternatives have forced OpenAI and Anthropic to compete on pricing for production workloads, not just raw performance. As developers realize they can route simpler tasks to efficient open models and reserve premium APIs only for complex work, established vendors have lost pricing leverage.