OpenAI Cuts GPT-5.6 Prices by Up to 80%, Three Weeks After Launch

On July 30, OpenAI cut prices on two of the three models in its GPT-5.6 family: Luna loses 80% of its price, Terra 20%, just three weeks after both models launched. The company attributes this to efficiency gains in its own infrastructure, but the timing lines up with mounting price pressure from Chinese AI labs.

The new rates

OpenAI attributes the cuts to progress in the model itself: end-to-end serving cost for GPT-5.6 dropped 20%, and token-generation efficiency improved by more than 15%. Sam Altman had already said plainly that cost is "a huge issue" for the company, which lines up with this decision rather than standing apart from it.

There's a second, less technical pressure behind this too: enterprise customers are increasingly reluctant to sign off on large AI spending without a clear picture of the return. For a company selling API access to its models, that translates directly into revenue: once a customer's procurement team starts counting token cost line by line, the price per million tokens stops being a technical footnote and becomes the actual argument in the sales conversation.

Pressure from China, less straightforward than it looks

On July 16, Moonshot AI released Kimi K3, an open-weight model with 2.8 trillion parameters that topped part of the Arena task-execution benchmarks and hit a 26.7% success rate on autonomous legal work testing, nearly double Anthropic's Claude Fable 5 result. Kimi K3 is also the most expensive model any Chinese lab has released to date: running the full Intelligence Index benchmark generated roughly 130 million output tokens and cost $2,709.75, averaging 94 cents per task. Moonshot defends that price directly, arguing that "state-of-the-art models deserve their own pricing power" instead of being automatically filed under "cheaper alternative."

For comparison, GPT-5.6 Sol averages $1.04 per task, and Claude Opus 4.8 costs $1.80. Kimi K3, despite being the most expensive model in Chinese AI history, still comes in cheaper than OpenAI's own flagship. Moonshot is meanwhile preparing a Hong Kong IPO at a $50 billion valuation, with CICC and Goldman Sachs advising, targeted for later this year.

Terra now undercuts Claude Sonnet 4.6

After the cut, Terra's new price ($2 per million input tokens, $12 for output) fell below Anthropic's Claude Sonnet 4.6, which costs $3 per million input tokens and $15 for output. A model priced above Anthropic's competition three weeks ago now costs less, with no change to the model itself beyond this one pricing decision.

What this says about GPT-5.6 itself

An 80% cut three weeks after launch is unusual in an industry where companies typically defend new-model pricing for months before discounts even enter the conversation. OpenAI frames this as pure technical optimization, but the pace of the change suggests something else too: that Luna and Terra's original pricing simply didn't survive contact with a real market, where enterprise buyers increasingly count the cost per token before approving an AI budget, and Chinese labs offer comparable reasoning and coding quality at a clearly lower rate.

SPAWNSY verdict