The price war reached the token. The residency surcharge did not move.

What happened on 22 September

OpenAI released GPT-6 Sol and GPT-6 Luna. On OpenAI's own pricing page Sol lists at $2.00 per million input tokens and $10.00 per million output tokens; GPT-5.6 Sol, the model it replaces, lists at a promotional $4.00 and $20.00. Luna lists at $0.10 and $0.50. Portfolio's table puts the previous Luna input price at $0.20. OpenAI's launch post refused our request (HTTP 403), so the price comparison rests on the pricing page and on Portfolio, which had read the announcement.

Anthropic released Claude Opus 5.5 the same day, at $4 and $20 per million tokens, which the launch page describes as 20% less than Opus 5. Cache reads, which Anthropic says make up the majority of agentic and coding work costs, fell to $0.20 per million, 60% less. Anthropic's own estimate is that at default settings the model costs 40% less than Opus 5 on typical workloads.

Portfolio's point was that the consumer chatbot is secondary. The stake is the price of running enterprise agents, which burn tokens by the million. On that reading the week was good news for every company that pays per token. We agree with the reading, and it applies to us.

Where we were wrong

On 3 September we wrote that the unit price can fall while the bill rises. On 4 September we wrote about a token price that stayed put while the bill grew. The thread running through both, and through a year of our conversations with Hungarian companies, was that public AI would not get cheaper.

Per token, it did. A 40% reduction on typical workloads is real money, and a company that pays OpenAI or Anthropic by the token will pay less on Monday than it did on Friday. We should say that plainly rather than reach for a caveat. The caveats exist, and we list them below, but they do not undo the fact.

There is a second correction, to ourselves. On 3 September we wrote that GPT-6 Astra had no EU Data Zone at launch. Microsoft's 22 September table now lists an EU Data Zone for Astra, at $12.00 and $60.00 against $10.00 and $50.00 Global. The option arrived. It arrived with a surcharge, and we guessed that part wrong too: we expected a premium similar to the US one, and it is double.

What did not move: the price of the location

Microsoft's Foundry announcement, by Naomi Moneypenny, carries a price table for all three GPT-6 models in three deployment types. For GPT-6 Sol, short context, Global Standard is $2.00 and $10.00, Data Zone Standard (US) is $2.20 and $11.00, and Data Zone Standard (EU) is $2.40 and $12.00. The table puts the US Data Zone 10% above Global. The footnote states the same 10% US premium in words, but only for the Provisioned Throughput and Priority Processing offers. The EU premium is not stated in words anywhere; it is arithmetic on the table, and it comes to 20% on every row of every model. Keeping the processing in Europe costs twice the premium of keeping it in the United States.

OpenAI's own pricing page says that regional processing endpoints, which is how it names data residency, are charged a 10% uplift for models released on or after 5 March 2026. It adds one sentence that matters more than the percentage: for GPT-6 Sol and Luna, EU data residency is available only with Standard processing. Our reading is that the discounted Batch and Flex tiers are for traffic that may be processed anywhere. If your data must stay in the EU, the sale price is not for you.

Anthropic's data-residency documentation is the shortest of the three. The inference geo parameter accepts two values, global and us. Pinning inference to the United States is billed at 1.1x. Workspace storage has one available geo, us. An EU company on Anthropic's first-party API can pay 10% extra to keep its data in the United States, and cannot choose Europe at all. Regional processing of Claude on the partner clouds is priced by the partners; Anthropic's own pricing page puts their regional endpoints at a 10% premium over global.

So the week's price cuts have a shape. The global price fell by 20 to 50 percent, and by more on Luna's output tokens. The residency premium stayed at 10 to 20 percent, or stayed unavailable. The token got cheaper. The location did not.

Then the seat stopped being the bill

On 25 September Microsoft relaunched Copilot as Home, Code and Autopilot, and split the way it is paid for. Jared Spataro's post says user subscription licences cover everyday work in Chat, Word, Excel, PowerPoint, Outlook and Teams at a fixed cost. Cowork, Code and Autopilot, what the post calls new long-running agentic capabilities, and frontier models like Astra and Fable, all run on usage-based billing. No prices were published. Microsoft's Partner Center notice of 16 September adds that from 2 November 2026 new Microsoft 365 Copilot Business licences bought through a Cloud Solution Provider include usage-based billing by default. Microsoft's message-centre notice MC1476200 adds that for new licences bought directly from Microsoft the change applies from 19 October.

We will not call this a surprise bill. Per the same notice the default configuration comes with spending limits, and Microsoft's post describes spending policies, API access for administrators and credit-request workflows, so the budget can be controlled. What changed is the shape of the invoice. The seat is a flat line. Agents and frontier models are a curve, and the curve rises exactly where every vendor is pushing adoption. A finance director who budgeted a licence per head in January now has a second line that is hard to forecast, because Microsoft gives what a Cowork task consumes only as rough planning estimates, and the post names no prices.