OpenAI just made its cheapest AI model far cheaper, and it did it barely three weeks after putting the model on sale. On Thursday the company cut the price of GPT-5.6 Luna, the fastest model in its lineup, by 80 percent, and trimmed its mid tier Terra by 20 percent. For the companies that build products on top of these models and pay by the token, the cost of running AI just dropped sharply, and the speed of the cut says the pricing power of the top labs is starting to slip.
The new numbers are steep. Luna now costs $0.20 per million input tokens and $1.20 per million output tokens, down from $1 and $6. Terra falls to $2 and $12 per million from $2.50 and $15. The flagship Sol model, the most capable of the three, keeps its price at $5 input and $30 output. In short, the two tiers most businesses would actually reach for got cheaper, while the premium model held its line.
OpenAI tied the cuts to plain engineering rather than a loss leading grab for market share. In an engineering post on July 29 it said a set of infrastructure improvements, including speculative decoding, finer grained load balancing, cache optimization, and custom GPU kernels, cut its serving costs by 20 percent and raised token generation efficiency by more than 15 percent. Cheaper to run means cheaper to sell. The company framed it as strategy, saying each generation of its models should accomplish “more work at a lower cost.”
The timing points to the real pressure. Chinese startup Moonshot AI released its openly available Kimi K3 earlier this month and beat some leading American models on benchmarks, Anthropic answered with Claude Opus 5 at half the price of its top public model, and Google rolled out three new models this month built to undercut rivals on cost. On Wednesday Microsoft chief Satya Nadella spent part of the company’s earnings call talking up cost efficient models. Everyone is racing the same direction, toward cheaper intelligence.
The comparison OpenAI clearly wants you to notice is with Anthropic. At the new price Luna undercuts Anthropic’s cheapest published model by roughly five times on input tokens, according to Reuters, and Terra now sits under the $3 per million that Anthropic charges for its mid tier Claude Sonnet. That is the whole point. OpenAI is not just lowering its own prices, it is trying to sit below the competition on the exact tiers where cost sensitive buyers make their call.
The cuts also land as companies rethink how much they spend on AI at all. After a stretch of urging staff to use it freely, many firms are now working to rein in ballooning bills. Amazon’s engineering organization moved to cap AI spending after cost overruns, and OpenAI itself shipped hard spend limits for API customers on July 22. A cheaper Luna is a direct answer to a customer who has started watching the meter.
There is a catch for the sellers. Analysts note that lower prices can lift usage but also squeeze the finances of OpenAI and Anthropic just as both head toward public offerings. And as Axios pointed out, cutting a model’s price three weeks after launch is unusual in an industry where discounts normally arrive months later. The message under the numbers is simple. The days when a brand new frontier model could hold a premium for long may be ending. For the wider view, see our coverage of the GPT-5.6 launch.
Your First 10 AI Skills
10 practical AI skills, copy-paste prompts and a 7-day plan to start using AI with confidence.
Download the guide →