OpenAI and Anthropic Cut Prices Below Their $10 Flagships

On September 22 Anthropic released Claude Opus 5.5 at prices 20% below the model it replaced, and within hours OpenAI released GPT-6 Sol at half the price of its predecessor. The coverage called it a price war, and it is one, just below each lab's flagship. Claude Fable 5.1 and GPT-6 Astra, the most expensive models each lab sells, kept their list rates. The discounts reach only work a company moves off that top rate, and US spending data shows buyers were already moving work to cheaper models.

The price war is real one step down

The strongest form of the price-war reading comes from Ara Kharazian, the economist who runs the spending data at Ramp, a US corporate-card company. He told Fortune on September 23 that the labs are fighting on two fronts: cheaper models that businesses are shifting toward, and outright price cuts on their most expensive ones. A buyer can draw a comfortable conclusion from that: cheaper AI will arrive without anyone having to decide anything.

The cuts behind it are real. Models are priced per million tokens, a token being roughly three-quarters of a word, with separate rates for what you send the model (input) and what it writes back (output). Opus 5.5 lists at $4 input and $20 output, per Anthropic's price list. GPT-6 Sol lists at $2 and $10, half the rate of GPT-5.6 Sol, and OpenAI told VentureBeat on September 22 that the new rates are permanent. It credits improvements in caching (reusing text the model has already processed) and in the cost of running the model. The saving holds up in use: in tests published the week of the launch, Artificial Analysis, an independent firm that benchmarks models, measured Sol's cost per task at 47% below its predecessor's. This week's Newsletter, Anthropic and OpenAI cut prices as access to AI gets contested, called price the easy part of the choice; the harder part is which of your work the new rates reach.

Buyers were already trading down

Ramp's September AI Index, published September 9 per Fortune and built from the spending of US businesses on its platform, puts the effective price those businesses paid per million tokens 41% below its March peak. Kharazian attributes the drop to a mix of lab price cuts and customers trading down to cheaper, simpler models. Ramp counts Opus, Sol and Fable together as top-tier models, and their share of tokens fell to 45% from a peak of 53% in August. Ramp says companies are setting company-wide defaults that send routine work to standard models, the cheaper group below.

The limits of the data: Ramp's sample is US businesses only, it does not say how much of the 41% came from lower prices and how much from the switch, and OpenAI gives cheaper serving, not customer behavior, as its reason for cutting. What the data does show is buyers moving work out of the top tier in the month before the September 22 cuts.

The flagship list rates held

Claude Fable 5.1, generally available since September 1, and GPT-6 Astra, launched two days later, both still list at $10 per million input tokens and $50 per million output. Fortune describes Sol as an offshoot of Astra. Anthropic did cut one flagship price: Fable 5.1 launched on September 1 with cache reads 75% cheaper than its predecessor's, at $0.25 per million tokens, per VentureBeat. That discount goes only to teams whose software is built to reuse the same text across requests.

Anthropic says Opus 5.5 performs at Fable's level on most work. That is the vendor's claim, and no independent test of it has been published yet. If it holds, a company that moves work from Fable to Opus 5.5 pays 60% less for each output token. The lab has published the cheaper option; the saving goes only to the customers who take it. Three weeks ago, in Meta and OpenAI Started Charging for AI by Who You Are, we argued that the list price had stopped being the price. This week shows the same pattern from the buyer's side: each discount arrives attached to an action the buyer has to take.

Write the default down

Two groups should be uneasy. The first is finance teams that treat vendor price cuts as their AI cost plan: the cuts reach only the work already running on the discounted models. The second is companies whose teams send every task to the flagship by habit. At list rates, an output token from GPT-6 Astra costs five times one from GPT-6 Sol, and one from Fable 5.1 costs two and a half times one from Opus 5.5.

Put the choice in writing. Sort your AI work into three or four task types, name the default model for each, and keep the flagship for work where a wrong answer is expensive: contracts, code that ships, anything a client sees unchecked. Then measure cost per finished task on 50 real tasks per type, retries included, because a cheaper model that needs a second attempt can cost more. Run the test again after every rate change. OpenAI changed its rates on July 30, August 21 and September 22, per its release notes.

The default decides the bill

The labs are competing hard on the models just below their best, and the flagship list rates have not moved since those flagships launched this month. For a buyer that splits the saving in two. The part the vendors hand out arrives on its own, and it reaches only the work already on the cheaper models. The other part depends on a decision most companies have never written down: which model does which job. Ramp's data says many US businesses have started making it. A company that has not will keep paying $50 per million output tokens for work the same labs now sell for $10 to $20, and no price war will change that number for it.