OpenAI Slashes Prices on Two GPT-5.6 Models as the AI Discount Wars Heat Up

OpenAI is cutting prices on two of its newest AI models just weeks after putting them on the market, a sign of how quickly the economics of the AI business are shifting.

The company announced Thursday that it's dropping the price of GPT-5.6 Terra, its mid-tier model, by 20%, to $2 per million input tokens and $12 per million output tokens. Luna, the fastest of the three GPT-5.6 models, gets a much steeper cut: 80% off, bringing it down to 20 cents per million input tokens and $1.20 per million output tokens. Sol, the top-tier model in the lineup and the most powerful of the three, keeps its current price.

Tokens are the chunks of text AI models read and generate, and per-token pricing is how most AI providers bill customers who build products on top of their models. For a company running millions of queries a month, small changes to that rate can swing costs by huge margins.

The GPT-5.6 family only launched about three weeks ago, which makes the timing of this price cut notable. It suggests OpenAI is responding less to its own product cycle and more to pressure building around it.

Enterprises are watching the bill more closely

That pressure is coming from customers who no longer treat AI spending as a blank check. In the early days after ChatGPT's 2022 debut, companies raced to roll the technology out across their workforces, often encouraging employees to use as much of it as possible without worrying much about the cost — a pattern that came to be known as "tokenmaxxing," according to CNBC.

That era is fading. As AI bills have climbed into the billions of dollars for some large customers, businesses have started demanding a clearer return on that investment before signing off on the priciest models. OpenAI's own language in its release Thursday reflects that shift. The company said its strategy is about "advancing both capability and efficiency so each generation of intelligence can accomplish more work at a lower cost."

The competitive squeeze

OpenAI isn't just responding to budget-conscious customers, though. It's also reacting to a fast-moving group of competitors racing to undercut it on price.

The clearest example came from Moonshot AI, a Chinese startup that released an open-weight model called Kimi K3 earlier this month. Open-weight models can be downloaded, modified, and run on a company's own infrastructure, which makes them a cheaper alternative to proprietary systems for many businesses. According to CNBC, Kimi K3 outperformed leading American models on some industry benchmarks, a result that got Silicon Valley's attention fast.

Anthropic responded almost immediately, releasing Claude Opus 5 and billing it as its best-performing and most cost-effective model for many use cases. It costs half as much as Claude Fable 5, the advanced model Anthropic put out in June, while matching it on coding and knowledge-work tasks, per the company.

Microsoft has been pushing the same message. CEO Satya Nadella repeatedly emphasized his company's cost-effective models during Microsoft's quarterly earnings call this past Wednesday, coming right after Microsoft launched what it called a cheap-but-capable cybersecurity model earlier in the week.

Google made a similar move this month, rolling out three new models built to beat competitors on price. The company says Gemini 3.6 Flash, the strongest of the three, costs less per task than Kimi K3 and other Chinese models.

What it signals

Taken together, the past month shows an industry pivoting hard toward cost as the main battleground, alongside raw capability. For years, the AI race was framed mostly around which model could do more. Now the question increasingly is which model can do enough for less.

That's a meaningful change for the companies buying this technology, who now have real leverage to shop between providers on price rather than defaulting to whichever lab has the flashiest benchmark. It's also a warning sign for margins across the industry, as providers keep cutting prices on models that were expensive to train in the first place.

For now, OpenAI still holds Sol back from any discount, keeping its most capable model at full price. That's a signal in itself: the company is willing to compete aggressively on cost for its everyday, high-volume models, while protecting the pricing of the model meant to showcase what it can do at the frontier. Whether that line holds as rivals keep pushing cheaper alternatives into the market is the thing to watch next.