Confirm the notice before you change anything
The premise of this guide is that fast and priority processing stop being available for EU-resident traffic on GPT-5.2, GPT-5.4 and GPT-5.4 mini on January 11. I could not confirm that from the pages I read. The OpenAI pages I opened returned navigation and no tier notice, and the policies index lists legal documents, not product changes.1 So treat the date, the model list and the post-cutoff behaviour as things to verify in the notice your account received, not as facts from this article.
That gap does not stop you from preparing. January 11 is 51 days after today, 21 November. Everything below works whichever way the notice words the outcome, because it is built around three possibilities you can test for, not one you have to guess.
Three questions decide the work:
- Which of your requests ask for the faster tier, directly or through a default?
- What does the provider do with such a request after the date?
- If it fails or is refused, where does your route send it next, and is that place still inside the EU?
The third question is the one compliance teams tend to find late, so most of this guide is about it.
Which requests lose the faster tier
Start with the population, not the code. A request is in scope when three things are true: it goes to an EU-resident project or deployment, it names one of the three models, and it was served at the faster tier.
The last condition is wider than "my code sets a field". OpenAI's Fast mode replaced Priority processing on 30 July 2026, requests tagged priority use Fast mode, and the tier can be set per request or as a project default.2 A service that never sets service_tier can still be served fast if its project default says so. Search for what was served, and for the project setting, as well as for the string in your repositories.
| Request class | How it ends up on the fast tier | How to find it |
|---|---|---|
| Explicit tier in code | service_tier set to priority or fast in the request body | Search clients, SDK wrappers and prompt-service configs; filter the request log by model |
| Project default | The OpenAI project is set to a faster default tier | Read the project's tier settings in the OpenAI Platform; OpenAI lets you choose which tiers each project may use2 |
| Gateway or SDK wrapper injection | A shared library adds the field for every call | Diff the outgoing body against what the application intended |
| Agent and batch jobs | An ETL or agent loop copied the interactive settings | Group requests by caller and look for sustained volume at the fast tier |
Then build a list of callers: for each, record the model, the volume per day, and whether a person is waiting on the answer. A person-facing chat path at the 95th percentile is a different problem from a nightly summariser that picked up a fast tier by accident. OpenAI itself advises against running large ETL or batch jobs in Fast mode.2 Those accidental users are the easy part: they move to standard and nobody notices.
Your own request log is the source of truth here. INFRO's request and usage records carry the model and an estimated provider cost per request, and the exports let you pull them for the audit.3 Use whatever your log records for the served tier, and where it does not record that, compare the estimated cost per token against the standard rate for the model.
What happens to a request that still asks for the tier
You do not yet know, so write down the three outcomes and handle each.
| Outcome after the date | What the application sees | What the route must do |
|---|---|---|
| Refused with an error | A 4xx before any output | Stop or retry at standard on the same EU target; never move to a non-EU target |
| Served at standard speed | A normal answer, slower, possibly billed differently | Nothing, but watch latency and the cost estimate |
| Field ignored | A normal answer; the field has no effect | Remove the field anyway so behaviour does not depend on it |
There is a documented precedent for the second outcome. Today, if Fast traffic ramps too quickly, OpenAI downgrades some requests to standard speed, bills them at standard rates and reports service_tier: "default" in the response.2 If the post-cutoff behaviour resembles that, the only evidence you get is that field and your latency. So record the served tier on every response now, before the date, so you have a baseline.
The refusal case needs the most care, because of how failover works. On INFRO, any upstream failure before the response starts, including a provider 4xx, 429, 5xx or timeout, moves to the next configured target.4 A refused tier is a provider 4xx. If the next target on the list is a global-region deployment or another provider outside the EU, your residency pin quietly becomes a fallback to somewhere else. The route's failure rules can change this per failure kind, to retry, move on at once, or stop,4 so decide per kind.
Which fallbacks keep EU residency
A fallback keeps EU residency only if the whole path does: the endpoint, the deployment type, the retention terms and the contract. Go through each candidate with the same questions.
| Candidate | Keeps EU residency? | What to verify |
|---|---|---|
| Same model, standard tier, same EU-resident OpenAI project | Expected to, if the notice leaves standard processing in the EU | The notice wording; the project's residency setting; your data processing terms |
| Azure OpenAI Data Zone Standard deployment | Processing stays within a defined geographic boundary5 | That the zone is the EU one; that your model and version exist there |
| Azure OpenAI Global Standard | Do not assume so: it is described as the broadest availability and lowest latency across regions5 | Treat as outside the boundary unless your agreement says otherwise |
| Azure provisioned deployment | Depends on type; regional provisioned is regional | Provisioned deployments are not auto-upgraded and need manual migration when a model retires5 |
| Another provider's EU endpoint | Only on that provider's own terms | Location and transfer terms, including retention |
| A third-party host of the same OpenAI ids | Often not available at all | Perplexity retired gpt-5.4, gpt-5.4-mini and gpt-5.2 from its Agent API on 24 October6 |
Two traps sit in this table. First, availability order: Microsoft says models reach Global Standard at launch, Data Zone deployments after the global provisioned type, and regional Standard last.5 A model version you rely on in the EU zone may exist later, or not at all, so check the exact version in the exact region. Second, the model itself may be on a retirement clock; on Azure, a retired model returns 410 Gone, and Standard, Data Zone Standard and Global Standard deployments are auto-upgraded on a rolling regional schedule.5 An auto-upgrade is a model change inside your residency boundary, and it still needs a re-test.
The gateway hop counts too. INFRO's EU compliance page says it does not currently provide guaranteed EU-only processing or data residency, that its gateway Worker sits near a Frankfurt database as a performance setting and not a guarantee, and that you should review the location and transfer terms of each provider, including every fallback target.7 If your policy requires every hop to be in the EU, a gateway on that footing is a decision for your compliance team, not something a route rule fixes.
Write the route rule
The shape of the rule is simple: match the traffic that was EU-pinned, list only EU-bounded targets, and make the failure behaviour explicit. INFRO routes over provider connections you own, with priority or weighted targets.4 If-rules send matching requests to their own targets, by metadata, key, project and similar conditions, which is how you separate the EU-pinned traffic from everything else.
Do not use routing.regions for this. INFRO's documentation says that field is recorded on the request and can conflict with an organization policy, but it does not region-constrain your connections and is not a data-residency control.4 The boundary is the target list.
What the rule should say, in words you can hand to a reviewer:
- Match requests tagged as EU-resident by project or metadata, so no caller has to opt in.
- Targets, in order: the EU-resident OpenAI connection at standard tier, then an EU data zone deployment of an equivalent model if you hold one. Nothing else.
- On a refusal that names the tier: stop, or retry once on the same target with the tier field removed. Do not move on.
- On a rate limit or a 5xx: retry the same target, then move to the second EU target.
- When no EU target remains: return the error to the caller. An error is a better outcome than a silent transfer.
Check the connection settings too. A connection can list alternate API hosts for the same key, tried first when the main host cannot be reached.4 Every listed host must be in the EU, or you have a hidden second path.
Finally, stop sending the field. Remove service_tier from clients and shared libraries, and remove the fast default from the project in the OpenAI Platform, so the behaviour on January 11 does not depend on how the provider handles a leftover parameter.
Test it before January 11
You need tests for three things: the tier is gone, the fallbacks stay inside the boundary, and the answers are still good enough at standard speed.
First, a replay set. Build a dataset from logged requests across the in-scope callers, with the long-prompt and tool-using ones included, and run it through a staging route that mirrors production. INFRO's datasets and experiments run a dataset through up to six variants, with a request count and estimated provider cost shown before the run and a spend cap it stops at.3 Variants here are the standard-tier target and the second EU target.
Second, a check for each failure mode. This harness sends one request per case and records what answered. It uses the OpenAI SDK against INFRO's OpenAI-compatible text endpoint,3 and the route slug and metadata key are placeholders for your own.
from openai import OpenAI
client = OpenAI(base_url="https://api.infro.io/v1", api_key="YOUR_STAGING_GATEWAY_KEY")
ALLOWED_MODELS = {"gpt-5.4", "gpt-5.4-mini"} # EU-bounded targets only; edit to yours
cases = [
{"name": "no tier field", "extra": {}},
{"name": "legacy priority tier", "extra": {"service_tier": "priority"}},
{"name": "fast tier", "extra": {"service_tier": "fast"}},
]
for case in cases:
try:
r = client.chat.completions.create(
model="eu-pinned-route", # your named route
messages=[{"role": "user", "content": "Reply with the word ok."}],
extra_body=case["extra"],
)
print(case["name"], "served by", r.model, "tier", getattr(r, "service_tier", None))
assert any(r.model.startswith(m) for m in ALLOWED_MODELS), "unexpected model"
except Exception as exc: # a refusal here is an acceptable outcome
print(case["name"], "refused:", type(exc).__name__)The script only proves the model name; it cannot prove where the request ran. For that, open the request traces: they carry the model, the connection that served it, the route and latency.3 For each case, confirm that the connection is one of your EU targets and that the attempts list contains no other.
Third, break the primary on purpose. Disable the EU-resident OpenAI connection in staging and repeat the cases. The expected result is either an answer from the second EU target or a clean error. Any answer from a connection outside your list is a failed test. The failover drill method covers how to stage this safely.
Fourth, measure latency at standard. The fast tier's pitch is output speed and consistent latency, so measure time to first token and total time at the 95th percentile on the standard target and set timeouts from those numbers; measuring time to first token covers the method. If a person-facing path no longer meets its target, the options are a smaller model, a shorter prompt or a streamed interface, not a route to a non-EU region.
A schedule for the 51 days
- Week of 24 November: confirm the notice text, affected models and the exact post-date behaviour with your OpenAI account contact.
- By 5 December: inventory in-scope callers from the request log and project settings; mark each interactive or background.
- By 12 December: remove the tier field from clients and the fast default from projects, starting with background callers.
- By 19 December: publish the EU-only route in staging and run the replay set and the failure cases.
- Early January: run the disabled-primary drill, record the traces as compliance evidence, and agree the error path with the product owners.
- Before 11 January: ship to production behind a small share of traffic, watch latency and the served tier, then go to full traffic.
What the gateway can and cannot carry
INFRO sits between your applications and the provider accounts you already hold, so the route rule, the per-request trace and the failure rules above can live in one place, on your own OpenAI and Azure connections. Model spend stays with your providers at the rates you negotiated. INFRO does not provide regional processing, so residency still depends on the connections you list and on your contracts. If you enter your negotiated rates, estimated cost uses them, which makes the move from fast to standard visible in your own numbers; see the routing documentation (opens in a new tab) and the EU compliance page (opens in a new tab) before you rely on either.
