Here’s the news, up front: On July 30, 2026, OpenAI slashed the price of its cheaper GPT-5.6 tiers. The Luna model dropped about 80% to $0.20 per million input tokens and $1.20 per million output tokens. The mid-tier Terra fell roughly 20% to $2 and $12. The flagship, Sol, didn’t move. If you build anything on AI, or you’re thinking about it, this GPT-5.6 price cut matters more than the headline suggests. Cheap tokens don’t just save money. They change the math on what’s economical to automate at all.
Let me explain why I care, and why you should too.
The numbers, and what they actually mean
Luna went from roughly $1.00 / $6.00 per million tokens down to $0.20 / $1.20. That’s an 80% cut on input and the same neighborhood on output, according to CNBC. Terra came down about 20%, from $2.50 / $15 to $2.00 / $12, per eWeek. OpenAI says the cuts come from efficiency gains across its models and inference systems, and the timing, three weeks after GPT-5.6 launched, lines up with mounting competition from rival Western labs and cheaper Chinese models, as VentureBeat reported.
A token is roughly three-quarters of a word. A million tokens is a fat novel. So we’re talking about processing a novel’s worth of text for twenty cents on the input side. That is absurdly cheap by the standard of even a year ago.
I haven’t run high-volume GPT-5.6 Luna workloads through the API myself, so I’m reading the pricing and the reporting here, not reporting from my own bill. But I’ve watched this curve long enough to know what it does.
Why the price of inference is the real story
The cost of “thinking” is collapsing, and it’s collapsing fast. This is the third or fourth meaningful price drop in this class of model in about a year. Each one quietly moves the line between “too expensive to bother” and “sure, automate it.”
Think about it in plain terms. Say you want AI to read every inbound support email, summarize it, tag it, and draft a reply. At $6 per million output tokens, running that across ten thousand emails a month adds up to real money, enough that you’d pause and build a business case. At $1.20, you stop doing the math. You just turn it on.
Cheap tokens don’t just lower your bill. They lower the bar for what’s worth automating in the first place.
That’s the shift. When inference was expensive, you reserved AI for high-value tasks. Now the economics invite you to point it at the boring, high-volume stuff, the ticket triage, the data cleanup, the “read this PDF and pull out the numbers” work that never justified a project before.
What this means if you run a small business
Your automations just got cheaper to run, and more of them just became worth building. If you’ve been holding off on an AI workflow because the token bill scared you, re-run the numbers. A monthly cost that looked like a phone bill might now look like a coffee.
A few concrete moves. Point a cheap model at the repetitive text work first: categorizing leads, drafting first-pass replies, summarizing long threads. Use the cheap tier (Luna-class) for volume and reserve the flagship for the handful of tasks that genuinely need the best reasoning. And wire it up through a tool you already understand rather than writing code. If you’re weighing your options there, I walked through the tradeoffs in my comparison of n8n, Zapier, and Make, and the pricing logic in that piece maps neatly onto this one: pick the tool whose cost curve matches your volume.
What this means if you run enterprise IT
Falling token prices are a double-edged budget line. On one side, your unit costs are dropping, which is great. On the other, cheaper inference means teams across your org will spin up far more AI calls, and aggregate spend can climb even as per-token cost falls. That’s the trap. Watch total consumption, not just the rate card.
This is also a moment to revisit model selection in your existing pipelines. If you provisioned everything on a flagship tier six months ago out of caution, a chunk of that traffic probably runs fine on a tier that now costs a fraction. Route by task. And don’t hard-code a single vendor’s model into your stack, because these prices are moving monthly and you want the freedom to chase them. The broader buildout I covered in my piece on the $206 billion enterprises are spending on AI agents this year only accelerates when the underlying tokens get cheaper.
Don’t overreact, either
Cheap does not mean free, and cheap does not mean good enough for everything. The Luna tier is the budget option for a reason. For anything where a wrong answer costs you a customer, a contract, or a compliance headache, test the cheaper model against your actual use case before you switch. Sometimes the flagship earns its price. Price is one input. Accuracy on your work is the one that pays the bills.
My take
This is the quiet trend that matters more than any single model launch. Models get smarter in headlines. Tokens get cheaper in the background, and that’s what actually decides which ideas make it out of your someday pile. Every one of these cuts turns a “not worth it” into a “why not.”
Your next step: pick one repetitive, text-heavy task in your business, the one you keep meaning to hand off. Estimate the tokens, run the new numbers at $0.20 / $1.20, and if it pencils out, build a small pilot this month. The tools are ready and the price finally is too.
News commentary by Brad Rowland — IT Infrastructure and Operations leader, automation builder, and AI implementer. Sources are linked inline.
