Article

Prepare OpenAI EU apps for January 11 tier changes

For EU app owners using GPT-5.2, GPT-5.4 or GPT-5.4 mini, check Fast defaults and compare Standard with supported model upgrades.

Editorial illustration for Prepare OpenAI EU apps for January 11 tier changes: a geometric block represents a model release. Not documentary evidence.

OpenAI’s November 18, 2026 notice sets January 11, 2027 as the retirement date for Fast/Priority processing on GPT-5.2, GPT-5.4 and GPT-5.4 mini when used with EU data residency. Affected app owners can keep the same model on Standard processing, or move to a supported replacement. The models themselves are not being shut down by this notice.

The first decision is whether your app can accept Standard responsiveness. If it can, changing the processing tier separates the deadline work from a larger model upgrade. If a supplier runs the app, ask it to confirm the model, regional endpoint and effective tier—not just the product name.

Find explicit requests and inherited Fast settings

Check the combination of model, region and processing tier. OpenAI’s Fast mode guide says both fast and priority select the same option on supported models. Fast can also be a project default, so searching source code for one parameter is not a complete inventory.

  • Find production calls using GPT-5.2, GPT-5.4 or GPT-5.4 mini. Record exact deployed IDs and any pinned snapshots; confirm ambiguous mappings rather than assuming every similarly named variant is affected.

  • Confirm the project’s regional configuration and endpoint. OpenAI documents eu.api.openai.com for Europe, covering the EEA and Switzerland, in its data-residency guide.

  • Inspect service_tier in request builders, shared wrappers, fallback routes and queued jobs. Also inspect Project Service Tier under Settings → Project → General. Requests can inherit Fast there even when application code never names it.

While reviewing request construction, check separately for saved prompt dependencies. The guide to moving OpenAI prompt objects into code covers that earlier November 30 deadline; changing tiers does not migrate those objects.

Keep the model and make Standard explicit

For a tier-only change, retain the supported model and EU route, and set service_tier to "default". Both the Responses reference and Chat Completions reference define this as Standard processing. The generation-request value is default, not standard.

For example, the relevant fields in a Responses request could be:

{
  "model": "gpt-5.4-mini",
  "input": "Summarize the supplied support request.",
  "service_tier": "default"
}

This is a request-body illustration, not a complete application. Keep your client pointed at the existing EU endpoint and preserve the app’s other required settings.

Do not simply delete the tier field. Omission behaves like auto, which follows the project setting. Explicit Standard on the affected requests gives a narrower change than changing a shared project’s default. If you do change that default, review every app that inherits it, and remove any explicit Fast overrides on affected routes.

Keep the regional boundary intact while changing speed. Residency coverage is specific to supported services and customer content; it does not extend to all system data or to third-party services merely because the OpenAI request uses an EU endpoint.

Treat a replacement model as a separate migration

OpenAI’s current regional support matrix lists GPT-6.2 Astra and GPT-6.2 Sol with EU Standard and Fast processing. Their Ultrafast option is US-only. Verify account access and the exact endpoint, model and tier combination before choosing either replacement.

A model-name change alone may not be enough. The GPT-6.2 migration guide specifies several compatibility changes:

  • Tools require the Responses API, even though text requests can also use Chat Completions. Replace unsupported none or minimal reasoning settings with low where applicable.

  • Remove unsupported sampling and log-probability options: temperature, top_p and top_logprobs; also remove logprobs from Chat Completions, or message.output_text.logprobs from Responses include.

  • Review prompt caching separately. For migrations from these older models, replace prompt_cache_retention with prompt_cache_options.ttl set to "30m". Put reusable instructions in a developer message at the start of the input for cache eligibility, and budget for cache writes.

Choose a replacement only after checking your actual tasks. Fast support establishes an available processing option, not a guaranteed response time or equivalent answer quality for your application.

Compare token costs before choosing an upgrade

Illustration using November 19 list prices: 10,000 short-context requests, each with 2,000 uncached input tokens and 500 billed output tokens, including reasoning. Hold usage constant: 20 million input and 5 million output tokens.

Totals below include the regional uplift: 10% for GPT-5.4, mini and GPT-6.2, but not GPT-5.2 under OpenAI’s release-date rule.

Model

EU Standard total

EU Fast total

GPT-5.2

$105.00

$210.00*

GPT-5.4

$137.50

$275.00*

GPT-5.4 mini

$41.25

$82.50*

GPT-6.2 Sol

$99.00

$198.00

Mini Standard: (20 × $0.75 + 5 × $4.50) × 1.10 = $41.25. Sol Fast: (20 × $4 + 5 × $20) × 1.10 = $198.

For identical usage, the tier-only switch halves these charges. Switching mini Fast to Sol Fast instead costs 2.4 times as much here—not a performance comparison.

Excluded: caching, tools, retries, discounts, taxes and engineering. Model changes can alter token counts; long-context rates differ. *Affected-model Fast totals describe the pre-retirement option only. Recheck prices before switching production traffic.

Check the result before moving all traffic

Use representative tasks to compare answer usefulness, tool results and any required structured output. Measure response times, slow requests and timeouts against what users need.

Inspect the returned service_tier as well as the value sent. OpenAI’s Fast mode documentation says affected older models report Fast usage as priority. It also documents a downgrade to Standard when traffic ramps too quickly; that is not a promise of automatic fallback after retirement.

If your release checks depend on OpenAI’s hosted Evals, address the separate November 30 transition using the Evals-to-Promptfoo migration guide. A model comparison needs a working evaluation process, not merely a new model ID.

Shift traffic in stages, then inspect delayed jobs and fallback paths again. Keep a supported EU configuration available for rollback; the retiring Fast combination will not be a durable rollback after the deadline.

Finish before January 11 rather than planning around a presumed midnight cutoff. The retirement notice gives no shutdown hour or timezone, exact post-cutoff error response, or automatic-migration guarantee.