EU buyers pay 10% more for GPT-6 Astra and cannot buy the fast tier at any price — and on Azure they cannot buy Astra in Europe at all

OpenAI’s GPT-6 Astra launched on 3 September with a benchmark table, a 2.5x price increase and an argument about why tokens are the wrong unit. The arithmetic in that argument is worth its own reckoning. This is a different problem, and it is not in the launch post at all.

It is in the model guide, in one sentence:

“GPT-6 Astra does not support service_tier: "fast" or service_tier: "priority" with EU data residency.”

That is not a price. It is a product boundary. And it means the Astra a European buyer can procure is a materially different product from the one on the launch chart — slower at the tail, with no escalation path, and about 10% more expensive for the privilege.

Three clouds, one number, and then a discontinuity

Start with what is genuinely settled, because it is the good news.

The cost of keeping inference inside a jurisdiction has converged. AWS put a public figure on it in August when Grok 4.6 arrived on Bedrock with a 10% data-residency premium. OpenAI’s pricing page states that regional processing endpoints carry a 10% uplift for models released on or after 5 March 2026. Microsoft’s own Astra listing prices the US Data Zone at precisely 10% over Global.

Three vendors arrived independently at the same surcharge. That is a market forming a price, and for anyone budgeting a regulated deployment it is a real improvement over the era when residency was a bespoke line in an enterprise contract.

Then look at the two bold rows. Those are not prices. They are the absence of a price, and no amount of budget converts them into one.

A capability tier is not a discount you declined

The distinction that matters here is between paying more and being unable to pay.

Fast mode is OpenAI’s latency product. It is reachable as either service_tier: "fast" or service_tier: "priority", it is marketed as up to 2.5x faster with more consistent latency, and it costs exactly double the standard rate. For a US-resident Astra deployment, that is a lever: when p95 degrades and the product team starts filing tickets, an engineer can flip a parameter and pay $20/$100 instead of $10/$50.

For an EU-resident deployment, that lever is not connected to anything. The parameter is rejected. There is no premium SKU, no enterprise escalation, no “contact your account director” path that the docs point at — the model guide simply says to use Standard processing.

This is the first time residency has cost a capability rather than a percentage, and it deserves to be read as a category change rather than a footnote. Every residency decision until now could be modelled as a cost: pick the jurisdiction, add 10%, move on. Astra introduces a residency decision that has to be modelled as a capability ceiling, which is a different kind of entry in a risk register and a different conversation with a product owner.

It is also worth being precise about who is actually hurt. If your EU workload is queued — batch summarisation, overnight agent runs, document processing, anything where a human is not waiting — Standard is what you would have chosen anyway and this restriction costs you nothing at all. The exposure is concentrated in interactive surfaces: assistants, voice, human-in-the-loop review, anything where tail latency is the product. Those teams have just lost their only knob.

The tier you can buy stopped being a promise

There is a second regression sitting alongside the first, and it applies globally rather than only in Europe.

OpenAI’s fast mode guide is explicit that fast mode for GPT-6 Astra does not include a latency SLA. For GPT-5.6 and earlier, the same guide describes fast mode and Scale Tier as receiving the same service-level treatment, with eligible enterprise agreements able to earn service credits when latency targets were not met.

So the ladder has shortened at both ends. The contractual rung became best-effort for everyone, and the best-effort rung disappeared entirely for EU-resident traffic. A US team paying 2x for Astra fast mode is buying a marketing claim about speed, not a commitment about it — which is a materially weaker purchase than the same parameter bought against Sol last month.

This is the same pattern visible when Anthropic withdrew the priority tier that promised uptime rather than keep selling a commitment it was missing. Frontier capability and contractual reliability are drifting apart across the whole category, and the newest model is consistently the one with the fewest guarantees attached. Anyone who watched four providers wobble inside the same three-hour window on 3 September already knows what the guarantees were worth in practice; the change here is that they are no longer being offered.

On Azure, the gap is not a tier — it is the whole model

The OpenAI-direct situation is a degraded Astra. The Microsoft Foundry situation is no Astra.

Microsoft’s own launch blog lists Astra in Global and US Data Zone deployments. There is no EU Data Zone, no other regional option, and no published timeline for one. Independent analysis of the Foundry listing reaches the same conclusion in blunter terms: no paid Astra deployment currently guarantees processing within the EU data boundary.

For a large number of European enterprises, Azure is the procurement path — the contracts are signed, the tenancy exists, security review has already run once. For those buyers, “GPT-6 Astra is generally available” is not true in any sense they can act on. It is available to a subsidiary in a different jurisdiction.

The pricing detail underneath makes the wait worse rather than better. Microsoft’s model deployment pricing update took effect on 1 September 2026 and applies its revised data zone premiums only to models launched on or after that date; customers who stay on their current models see no increase. Astra launched on 3 September. Whenever an EU zone arrives, it arrives priced under the new schedule — and Astra is the first flagship to sit on the wrong side of that line. Published estimates of the revised EU premium vary between roughly 9% and 20% across secondary coverage, and Microsoft has not posted Astra EU prices for the simple reason that there is nothing to price yet. Treat any specific figure circulating this week as an estimate, and plan the availability risk instead.

What this does to a model evaluation

The habit this should break is treating a model launch as a single event with a single date.

It is now a matrix. Jurisdiction × service tier × platform, and each cell has its own availability date, its own price and — as of this week — its own capability set. “GPT-6 Astra, GA, 3 September” describes exactly one cell of that matrix: global routing, US-resident, on OpenAI’s direct API. It is the cell the benchmarks were run in and the cell most launch coverage was written from. It may well not be the cell you deploy in.

That has a direct consequence for how benchmark tables should be read. Astra’s published results — the ones underpinning the split of frontier lineups into buyable and gated tiers — were produced without an EU residency constraint attached. That does not make them wrong; latency tier does not change model quality. It does make them incomplete as a procurement input for a team that will run the model on a tier the vendor benchmarked nothing on.

What to do about it

1. Classify the workload before you classify the model. Split EU traffic into latency-sensitive and queued. Only the first is affected. Teams that skip this step end up escalating a restriction that costs their actual workload nothing.
2. Re-run the price comparison at the tier you can actually buy. For EU-resident Astra that is Standard plus 10% — roughly $11/$55 short context. Compare that against Claude Fable 5.1, Gemini 3.8 Flash and the EU-domiciled alternatives on the same basis, not against Astra’s $10/$50 headline.
3. Check what your latency objective is written against. If an internal SLO or a customer commitment leans on OpenAI’s previous fast-mode service-level language, Astra does not inherit it. That is a document to update this week, not next quarter.
4. Put an Astra EU Data Zone date in the risk register, not the roadmap. On Azure there is no zone and no timeline. Anything that assumes one arrives before year end is an assumption, and it should be labelled as one.
5. Keep the routing layer between you and this. A gateway that can move EU traffic to a different provider without a code change is worth more this week than it was last week — the argument for neutral routing infrastructure has rarely been this concrete. Compare the shortlist properly: Claude against ChatGPT on residency terms, not just quality, and check the agent platforms you are building on for the same jurisdictional gaps.
6. Ask the vendor for the matrix in writing. Jurisdiction, tier, platform, date. If a sales team cannot fill that grid in, the answer to “is it GA in the EU” is no.

The bottom line

Nothing here is hidden. OpenAI documents the restriction in its model guide, prices the residency uplift on its pricing page, and states the missing SLA in its fast mode guide. Microsoft lists exactly two deployment types. Every fact in this piece comes from a vendor’s own documentation.

What none of them do is put it in the launch announcement. The 3 September changelog entry for Astra lists the model’s other constraints in detail — no none reasoning effort, no custom temperature or top_p, tool calling via the Responses API only — and says nothing about jurisdictions. It is the same silence that surrounds the three Responses API primitives shipped the same day, one of which decides whether an effort change costs $0.20 or $2.50 on a 200,000-token prefix. The EU AI Act transparency regime governs how AI output is disclosed to end users; it has nothing to say about whether a vendor’s launch post accurately describes where and how fast the model runs. That gap is a buyer’s problem to close, the same way the EEA consent-or-pay split on ChatGPT ads turned out to be.

So the honest summary of Astra’s first week in Europe is short. The surcharge is standard. The restriction is new. And the frontier model launch and the compliant frontier model launch are now two different events, separated by an interval nobody has announced.

Price the cell you will deploy in, not the one the benchmarks were run in.

Questions or corrections? Email Pick Right. Want the full list? See all news.