OpenAI Cuts GPT-5.6 Luna Pricing 80 Percent After Using Its Own Model to Rewrite Its Production Code

OpenAI dropped the price of its GPT-5.6 Luna model by 80 percent yesterday. Input tokens that cost 1 dollar per million are now 20 cents. Output pricing fell from 6 dollars to 1.20 per million. The mid tier model, Terra, is now priced at 2 dollars input and 12 dollars output per million, down from its earlier rate.

Both models launched three weeks ago. That's the entire window between release and an 80 percent price cut.

To put the token count in perspective, a million tokens works out to roughly 750,000 words in English, close to the first four Harry Potter books combined. Processing that much text now costs 20 cents.

The pricing number is the headline, but the method behind it is the more useful story for anyone building on these models.

OpenAI says it deployed its most capable model, Sol, inside Codex with a single task, optimize the code running its own production systems. Sol wrote parts of its own GPU kernels and decoding logic. That work cut running costs by 20 percent, and the savings passed directly to customers through the new pricing.

This wasn't a goodwill gesture. Two forces are pushing prices down across the industry.

The first is buyer behavior. CNBC reported that companies are scrutinizing AI spend more carefully than a year ago, and few are willing to pay premium pricing without a demonstrated return.

The second is competition from China. OpenRouter data shows American companies routing over a third of their workloads to Chinese models since February, spiking to 46 percent in some weeks, up from an 11 percent average the year before. Chinese models run several times cheaper for comparable tasks.

Get famous in your niche, build trust before the sales call, attract inbound leads, and create authoritative content that keeps working after every interview. PodPitch finds the right shows and handles personalized outreach automatically. Brands like Feastables use PodPitch. Only 25 demo spots are available this month.

Appear on 3,853,234 Podcasts

Against the rest of the market, Luna's new price is now near the bottom of the big lab tier. Claude Haiku 4.5 runs 1 dollar input and 5 dollars output per million tokens, still well above Luna. Google's Gemini 3.5 Flash-Lite costs around 2.80 dollars combined per million tokens. DeepSeek V4 Flash remains the cheapest capable model on the market at 0.14 dollars input and 0.28 dollars output, still undercutting Luna, but the gap between OpenAI's cheapest tier and the Chinese discount tier just got a lot smaller.

Sol's own price held steady at 5 dollars input and 30 dollars output, close to Claude Opus 4.8 and slightly above Gemini 3.1 Pro. OpenAI also added a new Fast mode for Sol, running two and a half times faster at double the cost. The pattern across the market is consistent now. Commodity tasks get cheaper every few weeks. Frontier capability stays priced at a premium.

For teams running AI tools or automations, the practical move is straightforward. Summarizing, tagging, classification, and support replies don't need a frontier model. Luna or a comparable budget tier handles those at a fraction of the cost. Save the expensive models for complex reasoning and coding work. That single routing decision can cut a monthly AI bill significantly.

Ten years ago, plenty of good ideas never became websites because someone couldn't scrape together hosting money. Now the price of the world's most capable technology is dropping toward the cost of a cup of tea, and the gap between the big labs and the cheapest challengers keeps closing.