Opus 5.5 vs GPT-6 Sol: Four Dollars Apart a Month

The short version: On Tuesday, September 22, Anthropic and OpenAI each cut the price of a flagship model, hours apart. Claude Opus 5.5 arrived at $4 per million input tokens and $20 per million output. GPT-6 Sol arrived at $2 and $10. Those are not simply two price cuts that happened to share a date. Opus 5.5 landed on precisely the rate card OpenAI retired that same morning, and both companies, independently, priced cached input at exactly twenty cents.

The number that matters to a small business is the one neither announcement published. A five-person shop running 880 customer inquiries a month through a drafted-reply workflow now pays about $4.40 in model costs on OpenAI’s top tier, or about $8.45 on Anthropic’s. On Monday, the same month on OpenAI’s flagship cost $8.80. The work got half as expensive overnight, and the distance between the two leading labs is now roughly four dollars a month.

What did each lab actually announce?

Anthropic released Claude Opus 5.5 on September 22, describing it as performing at the level of Claude Fable 5.1 on most work while costing 40 percent less to run than Opus 5. It generates output about 30 percent faster than Opus 5, and cache reads cost 60 percent less. It is available on the Claude Platform, Amazon Web Services, Google Cloud and Microsoft Azure.

OpenAI put GPT-6 Sol into its API the same day, alongside a smaller model called GPT-6 Luna, with a gradual rollout into ChatGPT Work and Codex for eligible paid plans. That last detail is worth reading twice if you are not a developer: the price cut is live in the API immediately, but arrival inside the ChatGPT product you actually log into is staged.

Here is what the two cards look like side by side, read from OpenAI’s pricing page and Anthropic’s announcement on September 23, 2026.

Read the first and third rows together. Anthropic’s newest flagship is priced at $4 and $20. That is, to the dollar, on both input and output, what OpenAI’s previous top model charged until Tuesday morning. One company’s new flagship costs exactly what the other company just walked away from.

What does a real month cost?

Rate cards are abstract, so here is a workload a real shop would recognize. An appliance repair business with five people handles 40 customer inquiries a day across calls, texts and web forms, 22 working days a month. That is 880 inquiries, each getting a drafted reply the owner reads and approves before it goes out.

Each request carries a standing context of roughly 2,000 tokens, which is the price list, the service area, the hours and the tone you want, and that part is cached because it never changes. Add about 300 tokens of the customer’s actual message, and about 400 tokens of drafted reply coming back.

That works out to 1.76 million cached tokens, 264,000 fresh input tokens and 352,000 output tokens a month. Run it through the cards:

The last inquiry of the day gets the same drafted reply as the first, and the whole month of them costs less than lunch. That is the part worth sitting with. The argument for pointing a model at this work was never about having fewer people in the shop. It was about the inquiries that used to sit unanswered until somebody got back off a job.

Note what the four rows show. The two frontier labs are four dollars apart on this job, while GPT-6 Luna does the same month for twenty-two cents. That is the 200 to 1 gap between flagship and cheap models we have written about before, and Tuesday did nothing to close it. If anything it widened.

Why did both prices land on twenty cents?

The cached input line is the one that quietly decides your bill, and it is the line neither announcement led with. Anthropic prices cache reads at $0.20 per million against $4 for fresh input, a twenty to one gap. OpenAI now prices cached input at $0.20 against $2, having halved it from the $0.40 that GPT-5.6 Sol charged.

Two competitors arriving at an identical cached-input price on the same day is not a coincidence of arithmetic. It is what a price war looks like when both sides can read each other’s published card. We wrote recently about a price war in which three frontier models launched in four days and none of them raised a price. This is the next move in that same war, and it is a sharper one: prices did not merely hold this time, they halved and converged.

For your business, that means the repeated part of your prompt is nearly free and the part that varies is not. A workflow that resends your whole price list fresh on every request pays twenty times more for it than one that caches it. That is a setup detail inside whatever tool you use, and a fair thing to ask your vendor about.

Token efficiency pulls the same way. Mario Rodriguez, GitHub’s chief product officer, said in Anthropic’s announcement that Opus 5.5 “used among the fewest tokens and steps” in testing. A model that reaches the same answer in fewer steps costs less than its rate card suggests, which is why ranking models on headline price alone misleads.

Will your software bill actually go down?

No. Not this month, and this is the part the coverage skipped.

You almost certainly do not buy tokens. You buy an AI answering service, or a scheduling assistant, or one of the tools inside the roughly $28 a month the median small business spends on AI software. Your vendor buys the tokens. When the model underneath them drops by half on a Tuesday, your invoice on Wednesday is identical.

The saving is real and it does reach you, just on a delay and through one of three doors. It reaches you at renewal, when your vendor reprices to stay competitive. It reaches you when a rival tool undercuts them and you move. Or it reaches you because you asked, which is the only one of the three you control.

So here is the useful thing to do with this news, and it takes five minutes. Find out which model sits underneath the AI tool you already pay for. Most vendors publish it, and the ones that do not will usually tell you if you ask. If it is a model that got cheaper this week, you have a reasonable and entirely friendly question for your next renewal.

Frequently Asked Questions

Is Claude Opus 5.5 or GPT-6 Sol better for a small business?

On cost, GPT-6 Sol is about half the price of Claude Opus 5.5 per token, which for a typical 880-inquiry month works out to roughly a four dollar difference. That gap is small enough that it should not decide anything on its own. Pick based on which one connects to the tools you already use, and which one drafts replies in a voice you are willing to put your name on.

How much does Claude Opus 5.5 cost?

Claude Opus 5.5 costs $4 per million input tokens and $20 per million output tokens, with cache reads at $0.20 per million and cache writes at $5, read from Anthropic’s announcement on September 23, 2026. Anthropic states it costs 40 percent less to run than Opus 5 on typical workloads.

Do I need to do anything to get the lower price?

If you call the API directly, you get the new pricing by pointing at the new model name. If you use an AI feature inside software you subscribe to, the lower cost goes to your vendor rather than to you, and your subscription price stays where it is until they reprice it or you ask them to.

Why is cached input so much cheaper than regular input?

Cached input is text the model has already processed and can reuse rather than read from scratch, so it costs the provider far less to serve. Both Anthropic and OpenAI now charge $0.20 per million for it. For a small business that sends the same price list or policy text with every request, caching is the single largest lever on the bill.

Something we keep wondering about: if the model underneath your AI tool just got 40 or 50 percent cheaper, who should raise that first, you or your vendor? We would genuinely like to hear how that conversation goes if you have it.