跪拜 Guibai
← All articles
AI Programming

OpenAI Slashes GPT-5.6 Luna Pricing by 80% as DeepSeek Prepares a Rate Hike

By ZJPRENO ·
Read original on juejin.cn ↗ Google Translate ↗ Alt translation

The cost of running high-performance AI models is bifurcating. A permanent 80% price cut from OpenAI makes a top-tier model accessible for broader production workloads, while DeepSeek's planned hike signals that compute constraints in China are now directly increasing costs for developers who built on its cheaper APIs.

Summary

A sharp pricing divergence is hitting the large language model API market. OpenAI permanently reduced the cost of its GPT-5.6 Luna model by 80%, dropping input to $0.20 and output to $1.20 per million tokens. The move is a direct result of inference optimization lowering operational costs, making a flagship-tier model cheaper than ever. Simultaneously, DeepSeek notified users of a planned, significant price hike across its API services, reversing its previous cost-leadership position.

The current numbers still favor DeepSeek for output costs in new queries, but the gap has narrowed dramatically. DeepSeek retains a massive advantage for context reuse through its cache-hit pricing, where input costs drop to near zero. However, the announced price increase threatens to flip the equation entirely, potentially making DeepSeek-Pro more expensive than GPT-5.6 Luna for standard calls.

For developers, the immediate takeaway is a new low-cost option for high-performance inference from OpenAI, while DeepSeek's peak-time doubling and looming price hike introduce new cost-planning variables. The contrasting strategies expose the underlying compute pressures facing domestic Chinese models versus the scaled efficiency gains of Western hyperscalers.

Takeaways
GPT-5.6 Luna input pricing is now $0.20 per million tokens and output is $1.20, an 80% permanent reduction effective July 30.
Mid-range Terra model pricing dropped 20%, while top-tier Sol pricing remained unchanged but gained a 2.5x faster mode at double the cost.
DeepSeek plans a significant overall API price increase soon, on top of existing peak-hour pricing that already doubles costs during Beijing daytime hours.
DeepSeek-Flash output remains cheaper than Luna at 2 RMB vs. 8.1 RMB per million tokens for new queries, but the gap has shrunk.
DeepSeek's cache-hit pricing drops input costs to 0.02–0.025 RMB, a massive advantage for multi-turn conversations and knowledge-base queries.
A post-hike scenario could push DeepSeek-Pro costs above GPT-5.6 Luna for standard calls, reversing the current cost hierarchy.
Conclusions

OpenAI's ability to permanently cut prices by 80% without a short-term promotion suggests inference optimization has structurally lowered its unit economics, not just marketing.

The price cut positions Luna's benchmark scores at parity with the old, far more expensive GPT-5.4, effectively delivering a generational price-performance leap in one move.

DeepSeek's peak-time pricing and planned hike expose a hard ceiling for domestic Chinese AI providers: GPU scarcity and high compute costs are now directly passed to developers.

The recommendation to avoid DeepSeek during working hours turns API pricing into a scheduling problem, adding operational complexity that a flat-rate global provider avoids.

Concepts & terms
Cache Hit Pricing
A billing model where repeated or identical context in API calls is charged at a drastically lower rate because the model reuses previously computed results instead of reprocessing the input from scratch.
Peak/Off-Peak Time-of-Use Pricing
A pricing strategy where API call costs vary by time of day, doubling during high-demand hours to incentivize off-peak usage and manage server load.
Source: juejin.cn ↗ Google Translate ↗ Backup ↗