OpenAI launched two cheaper members of the GPT-6 family on September 22, pricing GPT-6 Sol at $2 per million input tokens and $10 per million output, and GPT-6 Luna at $0.10 and $0.50 - half the rate of the GPT-5.6 tier they replace, by the company's own framing. Hours earlier, Anthropic released Claude Opus 5.5 at $4 and $20 per million tokens against Opus 5's $5 and $25, and said it costs about 40 percent less to run on typical workloads.
Neither announcement was framed as a cut. Both were framed as efficiency. OpenAI attributed the lower prices to improved caching and inference work, and said the savings were being passed on. Anthropic reported that Opus 5.5 used roughly half the tokens and made about 40 percent fewer calls than Opus 5 on a public benchmark of real command-line tasks, and that it beats Claude Fable 5.1 on key agentic benchmarks - including Terminal-Bench 4.0 - while costing less per million tokens. The two companies described a technical achievement and then lowered the number attached to it.
The market read it as a single event anyway. CNBC and SiliconANGLE described Anthropic's release and OpenAI's response as a same-day back and forth; the Financial Times noted that U.S. frontier model prices have fallen sharply since mid-July under pressure from Chinese competitors. Fortune quoted Ramp lead economist Ara Kharazian calling it an all-out price war - one that lowers the price of AI and, with it, the labs' ability to profit from what they sell. Ramp's own index put the effective price per million tokens down about 41 percent from a March peak of $1.15 to $0.68.
What pacing looks like from the billing side
This is the ninth day of a particular sequence. On September 12 Anthropic's chief executive published a case for pacing the frontier, arguing that capability gains should slow until alignment, security, and evaluation work catch up. Within a week, OpenAI's Sam Altman told staff he was open to slowing down on the cutting edge and hoped other labs would follow. Both men were then named as defendants in a class-action antitrust complaint in the Northern District of California, filed by subscribers who argue that coordination of that kind held back products they paid for.
That complaint is about restraint of trade. What the labs actually shipped looks more like restraint of margin. If capability growth is deliberately paced while competition continues, the axis of competition moves - and it moves to exactly the thing a paying customer can measure without a benchmark suite. Opus 5.5 is the first Anthropic release since the slowdown call, and its headline number is a price. GPT-6 Sol and Luna are OpenAI's answer to it, and their headline numbers are also prices.
The cheaper tier is the product
There is a second reading, less diplomatic and probably more accurate. Luna at ten cents per million input tokens is not a frontier model with a discount. It is a different product for a different bill, sold to people who were never going to pay frontier rates for summarizing a document. The price war is happening at the bottom of the market, where volume lives, and the announcement that matters to most developers is not that intelligence got cheaper but that a usable amount of it now costs a rounding error.
The capability story has not gone away. It has simply stopped being the headline for one news cycle, replaced by the only figure that appears on an invoice. Two labs spent the summer arguing that the frontier should not move too fast. On September 22 they both published evidence that it is moving - sideways, toward cost.
What a Price War Costs the Seller
A falling price is unambiguously good for the buyer and unambiguously bad for the seller, and both of these companies are sellers with obligations. Anthropic arranged a fifteen-billion-dollar revolving credit facility in early September and was reported to be weighing a public listing around mid-October. OpenAI has committed to compute contracts whose scale assumes it can charge for what runs on them. Halving the rate is not a rounding error against those commitments; it is a direct reduction in the revenue that services them. The industry has spent two years explaining that compute costs money and that whoever holds the compute holds the position. On September 22 two of the largest model vendors cut the price of the output. Neither said what it does to the arithmetic, and the question is now open in a way it was not a week ago.
The Takeaways
- OpenAI priced GPT-6 Sol at $2/$10 per million tokens and GPT-6 Luna at $0.10/$0.50, about half the GPT-5.6 rate they replace.
- Anthropic says Claude Opus 5.5 costs about 40 percent less to run on typical workloads, at $4/$20 per million tokens against Opus 5's $5/$25.
- Both companies framed the reductions as efficiency gains rather than cuts; the market read the same-day timing as a price war.
- The Financial Times links the decline to pressure from Chinese competitors, with U.S. frontier prices falling sharply since mid-July.
