OpenAI Drops Prices in Half — But the Timing Says Everything About the AI Price War

OpenAI has a habit of releasing big news right on the heels of its biggest rival, and this week it happened again.

On September 22, 2026, the company launched two new models, GPT-6 Sol and GPT-6 Luna, expanding the GPT-6 family that started with the flagship model Astra earlier this month. Here's the interesting part though — this launch came almost exactly 90 minutes after Anthropic released an updated version of its own model, Claude Opus 5.5, with lower prices of its own. Whether that timing was a coincidence or a calculated move, it's hard to ignore what it says about how intense the rivalry between these two companies has become.

What Are Sol and Luna, Exactly?

Think of OpenAI's lineup like a three-tier system. Astra is the flagship — the most powerful model, meant for the hardest, most important work. Sol sits a level below, built for everyday complex tasks like coding and business workflows. Luna is the budget option, meant for simpler high-volume jobs like summarizing documents or answering quick questions.

Sol and Luna aren't brand new concepts — earlier versions existed under the GPT-5.6 name. What's changed now is that OpenAI trained these new versions using the same techniques that made Astra so capable, then handed them a 50% price cut.

To put it simply:

• GPT-6 Sol now costs $2 per million input tokens and $10 per million output tokens — half of what it cost before.
• GPT-6 Luna costs just $0.10 for input and $0.50 for output, making it one of the cheapest capable AI models on the market right now.

For comparison, Sol now costs about a fifth of what Astra costs to run, and Luna is priced in the same range as many free, open-source models that companies host themselves.

Interesting Twist: No "Terra" This Time

If you were paying attention to OpenAI's earlier lineup, you might remember there was a third model in this family called Terra, sitting in the middle between Sol and Luna. This time, Terra didn't get an upgrade. According to OpenAI's own performance numbers, Luna at its highest setting now does roughly what Terra used to do, and Sol is priced even lower than Terra was a few months ago. So the lineup has effectively been simplified from three tiers down to two.

The Benchmark Battle

OpenAI didn't just talk about price — it also published a long list of comparisons against competing models, mostly Anthropic's Claude Opus 5 and Claude Fable 5.1.

On a workflow test called AutomationBench, which checks how well an AI can handle real business tasks like sales and HR work, Sol scored higher than Opus 5 while costing about 11 times less per task. On a coding benchmark called DeepSWE, Sol came within about one percentage point of Fable 5's top score, but at roughly a fifth of the cost.

OpenAI also claims Sol now makes about half as many factual mistakes as its previous version, based on real conversations where users had flagged errors.

That said, it's worth mentioning a small catch here — on a couple of these same tests, when compared to the older GPT-5.6 Sol's best possible scores (numbers OpenAI published when it launched Astra), the new GPT-6 Sol actually scores slightly lower. So while the price-to-performance ratio has clearly improved, raw peak performance on a couple of tests didn't necessarily go up.

A Smarter Way to Save on Long Conversations

Beyond just cutting prices, OpenAI also improved something called "prompt caching" — basically, a way for AI systems to remember and reuse earlier parts of a conversation instead of reprocessing everything from scratch every single time. This matters a lot for AI agents doing long, multi-step tasks.

Developers can now switch an AI's reasoning effort up or down, or turn tools on and off, in the middle of a task without losing that cached memory — something that wasn't possible before. OpenAI says this alone has cut the amount of repeated processing by more than half across billions of requests, based on data shared by GitHub, which uses these models to power Copilot.

Where Can You Actually Use It?

Right now, GPT-6 Sol and Luna are available inside ChatGPT Work and Codex for anyone on a paid plan — Plus, Pro, Business, Enterprise, or Edu accounts. Free users only get access to Luna, and only through the desktop app for now. Developers can also access both models directly through OpenAI's API.

One thing worth noting — these models are not yet available in regular ChatGPT chat, the version most casual users interact with daily. OpenAI says that's rolling out gradually throughout the day.

The Bigger Picture

This release says a lot about where the AI industry currently stands. It's no longer just about who can build the smartest model — it's increasingly about who can deliver strong performance at the lowest possible cost. Every major lab now publishes its results as "score per dollar" instead of just raw performance numbers, and this launch leaned heavily into that framing.

It's also a reminder of just how closely OpenAI and Anthropic are watching each other. A 90-minute gap between two competing product launches isn't something that happens by accident in an industry this competitive — and it likely won't be the last time either company times a release around the other's announcement.