"Default Price" Matters More Than "Half Off"
The two sides of the comparison are not the same kind of thing: the GPT-5.6 tier's pricing was promotional, while per an OpenAI spokesperson, Sol's and Luna's rates are the default. For anyone budgeting, that distinction matters more than the size of the cut. **A promotional price means marking an expiry date on the calendar and preparing to renegotiate then; a default price means the anchor moved.** The first is a debt coming due. The second is a number you can put in a three-year cost model. This site covered the mirror image of this on September 1: Claude Sonnet 5 launched with $2/$10 labeled promotional through August 31, with $3/$15 due to take effect September 1, and Anthropic ultimately canceled the increase, making $2/$10 the standard price. The observation then was that many third-party pricing guides still listed the old numbers — **a change in the nature of a price is easier to lose in transmission than a change in the number**, and that is the same point today. While we are here, a correction that often gets lost in retelling: Luna's output price fell from $1.20 to $0.50, about 58%, not 50%. Only the input tier is exactly half.
Three Companies Moved on Price Within 48 Hours, and Not on Benchmark Scores
The timeline is worth recording: per TechCrunch, Anthropic's Opus 5.5 shipped roughly 90 minutes before OpenAI's announcement; xAI's Grok 4.7 landed inside the same 48 hours; and Google's promotional pricing on Gemini 3.8 Flash runs through December. What best captures the shape of this competition is an almost coincidental result: **Sol's price is exactly half of Opus 5.5's on both measures** ($2 against $4, $10 against $20). Sol's pricing also matches Claude Sonnet 5. So on a single day, two companies placed their mid-tier workhorses at a 2x price gap rather than jockeying over a benchmark score. This site has a separate piece today on Opus 5.5, and reading them together makes the picture clearer: inference pricing is now in direct competition, and each company is using a different instrument. OpenAI opened cheaper tiers and declared the rates default. Anthropic cut the flagship 20%, cut cache reads 60%, and added subscription headroom. **Where the cut happens determines who benefits:** a new tier helps those willing to switch models, a cache cut helps those re-reading long contexts, and subscription headroom helps those who never touch the API.
The Most Useful Number Comes From Real Production Traffic
One item in this announcement is worth more than any benchmark: OpenAI says its caching improvements cut the volume of tokens needing fresh processing by more than half, measured across billions of requests flowing through GitHub Copilot. **That is production traffic, not an evaluation set.** Layer the 90% discount on cached input reads on top, and for applications with heavily repeated context — coding assistants, support, retrieval-augmented systems — the improvement in an actual bill may far exceed the 50% on the price list. By the same logic, if your calls share almost no reusable prefix, the price list's half is all you get. The two performance claims, by contrast, need reading on their terms. **OSWorld 2.0's 60.5% against 60.3%** has Sol at xhigh and Opus 5 at medium. That is not a like-for-like comparison; it is a pairing the vendor chose. "About 80% lower cost per task" is a legitimate commercial metric, but it answers how much you pay to reach that score, not which model is stronger. **Luna at higher effort matching GPT-5.6 Sol at about one-hundredth of the cost** would, if it holds, be an order-of-magnitude change for high-volume classification, routing and extraction. But it is an internal test — and those are exactly the tasks easiest to verify yourself. Run a batch of your own samples and you will have an answer within the hour.
API-Only, No Weights: a Line Worth Recording
Sol and Luna are API-only models with no weights to self-host. Set that beside two stories from September 21 and the week completes a set of three postures: **Qwen-Image-2.1 shipped weights but pulled its license back to research-only; StepFun's Step 5 Preview promises weights on October 15; and this GPT-6 tier offers no weights at all.** For anyone who must self-host — data that cannot leave a jurisdiction, offline deployment, or pinning a specific version long-term — the price cut is irrelevant, because GPT-6 is not on that path. One last item on the calendar: this launch lands ahead of DevDay on September 29. Cutting prices to a new default a week before a developer conference usually means there is something else to announce there.
via: The New Stack, The Next Web, MarkTechPost, Quartz