On this page · 12 sections
- What OpenAI announced on 18 November
- Who is affected
- Find the traffic before you choose
- Option A: stay on GPT-5.x with Standard processing
- Option B: move to GPT-6.2 and keep Fast in the EU
- The cost model, in dollars, euros and rupees
- A 52-day plan to 11 January
- Which option suits which workload
- Data protection angle
- India-specific considerations
- How eCorpIT can help
- References
Summary. On 18 November 2026 OpenAI added an entry to its deprecations page: from 11 January 2027 it stops offering Fast and Priority processing for GPT-5.2, GPT-5.4 and GPT-5.4 mini when they run under EU data residency. That is 54 days of notice, and 52 days remain from today, 20 November 2026. The models are not being retired. OpenAI names two routes: keep the same models on Standard processing (service_tier: default), or move to the GPT-6.2 family, launched on 17 November 2026, whose Astra and Sol models support EU residency with Standard, Fast and Batch. The money matters. For a workload of 400 million input and 40 million output tokens a month on GPT-5.4, EU Fast costs about $3,520 (€3,202); EU Standard costs $1,760; GPT-6.2 Sol costs $1,320 on EU Standard and $2,640 on EU Fast, all including the 10% residency uplift. Our verdict: GPT-5.2 and GPT-5.4 teams that need speed should test GPT-6.2 Sol on Fast now; GPT-5.4 mini teams should first check whether Standard latency is good enough, because no GPT-6.2 model comes close to the mini's price.
About this article. This is for engineering leads, platform teams and data protection owners who call OpenAI from an EU-residency project, whether in Europe or from India serving European customers. On 20 November 2026 we checked every date, price, parameter and support statement against OpenAI's deprecations page, changelog, pricing page, Fast mode, data residency and background mode guides, the GPT-6.2 migration guide, five model pages and the Responses API reference, plus the European Central Bank's reference rates of 19 November 2026. Prices and support tables change, so re-check before you commit budget. This is not legal advice, and no vendor paid for inclusion. Researched with AI assistance and checked against the primary sources listed in the references.
What OpenAI announced on 18 November
The entry sits in the feature and endpoint deprecations table, not the model table, which matters for how you read it.
| Item | What the deprecations page says |
|---|---|
| Affected combination | GPT-5.2, GPT-5.4 and GPT-5.4 mini with EU data residency and service_tier fast or priority |
| Status | Deprecated |
| Notice date | 18 November 2026 |
| Shutdown date | 11 January 2027 |
| Replacement 1 | The same models on EU data residency with Standard processing (service_tier default) |
| Replacement 2 | The GPT-6.2 model family |
Four points follow from the table and the surrounding docs.
Only one combination ends. GPT-5.2, GPT-5.4 and GPT-5.4 mini stay available on Standard in the EU, and the notice does not touch Fast outside EU residency. The EU support table still lists the dated snapshots gpt-5.2-2025-12-11, gpt-5.4-2026-03-05 and gpt-5.4-mini-2026-03-17 for the Responses endpoint.
Fast and Priority are the same thing. OpenAI renamed Priority processing to Fast mode on 30 July 2026, and both service_tier: fast and service_tier: priority select it. For GPT-5.6 and earlier models the response reports priority whichever value you sent, so a log search for only one of the two strings will miss traffic.
The notice is short by OpenAI's standards for models. OpenAI promises at least six months before it retires a generally available model. That policy is written for model retirement; this change removes a processing tier in one region while the models live on, and it arrived with under eight weeks to go.
OpenAI has not said what a non-compliant request does after 11 January. The API reference says the response body reports the tier that actually served a request, which may differ from the one you asked for. Whether an EU request that still asks for Fast after the deadline will fail or be served at Standard is not documented. Do not plan around either outcome; change the request.
Who is affected
You are affected if all three of these hold:
- Your traffic runs on EU data residency, sent to
eu.api.openai.com. OpenAI's current code examples show that EU processing needs an API key from an EU-region project, even where a Global project uses per-request regional processing for other regions.
- The model is
gpt-5.2,gpt-5.4orgpt-5.4-mini, including their dated snapshots.
- Requests run in Fast mode, either because they set
service_tiertofastorpriority, or because the project's default service tier is Fast.
The third case is easy to miss. The Fast mode guide explains that a project's service tier can be set to Fast in project settings, and requests that omit service_tier then run in Fast mode by default. The API reference says the default auto value takes whatever the project setting is. OpenAI's notice names the request parameter, not the project setting, so treat an EU project with a Fast default as in scope and check it.
You are not affected if your EU traffic uses other models (the notice names only these three), if your GPT-5.x traffic already runs on Standard or Batch, or if your Fast traffic is not on an EU-residency route. Our briefing on per-request regional processing explains how prefixed domains and Global projects interact.
Find the traffic before you choose
Three checks give you an inventory in an afternoon.
Usage dashboard. OpenAI's Fast mode guide says you can group usage by service tier in the dashboard; filter to your EU projects and look for the priority tier against the three model names.
Code and config search. Search repositories, gateway configs and environment files for service_tier, fast, priority and eu.api.openai.com. Many teams set the tier in an AI gateway rather than in application code; if you route through one, the change belongs there. Our AI gateway comparison covers how LiteLLM, Cloudflare and others handle routing and configuration.
Response logging. Log the service_tier field returned on every response. It shows what actually served the call, which is the only reliable answer when a project default or a gateway rewrites the request.
A short check you can run in CI or as a scheduled job:
import re
AFFECTED = re.compile(r'^(gpt-5\.2|gpt-5\.4|gpt-5\.4-mini)(-\d{4}-\d{2}-\d{2})?$')
def check_request(model: str, service_tier: str | None, base_url: str) -> str | None:
is_eu = 'eu.api.openai.com' in base_url
if is_eu and AFFECTED.match(model) and service_tier in ('fast', 'priority', None, 'auto'):
return f'{model} on EU residency with tier {service_tier}: set Standard or move to GPT-6.2 before 11 Jan 2027'
return None
The pattern matches the three model aliases and their dated snapshots only, so gpt-5.4-nano and gpt-5.4-pro are not flagged. The None and auto cases are flagged because they inherit the project default, which may be Fast.
Option A: stay on GPT-5.x with Standard processing
This is the smallest code change: set service_tier to default on every affected request, and set the EU project's default tier back to Standard. OpenAI names it as a supported route.
What you give up is latency. OpenAI says Fast mode delivers up to 2.5 times faster speeds and more consistent latency, aimed at user-facing applications with regular traffic. How much of that you lose depends on the model and the request, so measure time to first token and tokens per second on your own prompts rather than assume the headline figure.
What you gain is a halved bill. Fast costs exactly twice the Standard rate for all three models on OpenAI's pricing page.
| Model | Standard input per 1M | Standard output per 1M | Fast input per 1M | Fast output per 1M | EU residency uplift |
|---|---|---|---|---|---|
| GPT-5.2 | $1.75 | $14.00 | $3.50 | $28.00 | None listed |
| GPT-5.4 (under 272K) | $2.50 | $15.00 | $5.00 | $30.00 | 10% |
| GPT-5.4 mini | $0.75 | $4.50 | $1.50 | $9.00 | 10% |
The uplift column comes from the pricing page and model pages: regional processing endpoints carry a 10% uplift for models released on or after 5 March 2026, and the GPT-5.4 and GPT-5.4 mini pages state it directly. GPT-5.2's snapshot is dated 11 December 2025, before that cut-off, and its page lists no uplift.
Option A suits back-office work that was put on Fast out of habit, internal tools where a second of extra wait does not matter, and any team that wants to change nothing else before January. It is also a holding position: the GPT-5.2 model page already recommends GPT-6.2 Astra, so a model move is coming either way.
Option B: move to GPT-6.2 and keep Fast in the EU
The GPT-6.2 family has two models. OpenAI positions GPT-6.2 Astra for the most demanding tasks and GPT-6.2 Sol for complex work at lower cost. Both support US and EU data residency with Standard, Fast and Batch processing; Ultrafast is US-only.
| Property | GPT-6.2 Sol | GPT-6.2 Astra |
|---|---|---|
| Standard input and output per 1M (short context) | $2 and $10 | $8 and $40 |
| Fast input and output per 1M (short context) | $4 and $20 | $16 and $80 |
| Cached input per 1M (Standard) | $0.10 | $0.40 |
| Context window | 1,050,000 tokens | 1,050,000 tokens |
| Maximum input tokens | 922,000 | 922,000 |
| Reasoning effort none | Rejected | Rejected |
| EU residency with Fast | Yes | Yes |
The figures come from the two model pages and the pricing page. Above 272,000 input tokens, input and cache rates double and output rises to 1.5 times the short-context rate for the whole request.
For most GPT-5.2 and GPT-5.4 workloads, Sol is the natural target: its Standard rates sit below GPT-5.4's, and its Fast rates sit below GPT-5.4 Fast. Astra costs four times Sol and is worth testing only where quality on hard tasks justifies it.
The data residency guide lists three more models that keep Fast in the EU: GPT-6.1 Sol, GPT-6.1 Terra and GPT-6 Luna. GPT-6 Astra and GPT-6 Sol do not. The deprecation notice names only the GPT-6.2 family, but a GPT-5.4 mini workload that genuinely needs EU Fast may find GPT-6.1 Terra a closer price match: its Fast rates of $2 and $10 per million are half of Sol's. Our cost comparison of GPT-6.1 Sol, Terra and Luna works through those tiers.
What breaks when you change the model string
GPT-6.2 is not a drop-in swap. OpenAI's migration guide and model pages list these changes:
- Reasoning effort. GPT-5.2, GPT-5.4 and GPT-5.4 mini default to
none; GPT-6.2 rejectsnone. OpenAI suggestslowif you usednoneorminimal, andmediumfor a new integration. Expect more reasoning tokens, and so more output cost, than a GPT-5.x call atnone.
- Sampling parameters. Remove
temperature,top_p,logprobsandtop_logprobsfrom Chat Completions requests, andtemperature,top_pandtop_logprobsfrom Responses requests.
- Tool calling. GPT-6.2 supports Chat Completions, but tool calling needs the Responses API.
- Prompt caching. GPT-6.2 appends top-level
instructionsafterinput, so move stable instructions into a developer message at the start ofinputto keep them cacheable. Teams coming from GPT-5.5 or earlier replaceprompt_cache_retentionwithprompt_cache_options.ttl.
- Prefix edits. The Responses API drops reasoning items after edits to earlier conversation history or certain model changes, which can lower quality and raise token use.
- No Flex. GPT-6.2 does not support
service_tier: flex. For slow work, the Responses API offers background requests with two-hour and eight-hour completion windows at 30% and 40% below Standard, in public beta.
Each item is a code change and a test, which is why the move to GPT-6.2 needs to start this month, not in the first week of January.
The cost model, in dollars, euros and rupees
The table prices one example workload: 400 million uncached input tokens and 40 million output tokens a month, short context, on an EU-residency endpoint. We applied the 10% uplift to every model released on or after 5 March 2026, and none to GPT-5.2. The first row works out as (400 x $5.50) + (40 x $33.00) = $2,200 + $1,320 = $3,520, where $5.50 and $33.00 are the Fast rates of $5 and $30 plus 10%.
| Route | EU input per 1M | EU output per 1M | Monthly US dollars | Monthly euros | Monthly rupees |
|---|---|---|---|---|---|
| GPT-5.4 on Fast (ends 11 Jan) | $5.50 | $33.00 | 3,520 | 3,202 | 3,28,416 |
| GPT-5.4 on Standard | $2.75 | $16.50 | 1,760 | 1,601 | 1,64,208 |
| GPT-6.2 Sol on Fast | $4.40 | $22.00 | 2,640 | 2,402 | 2,46,312 |
| GPT-6.2 Sol on Standard | $2.20 | $11.00 | 1,320 | 1,201 | 1,23,156 |
| GPT-6.2 Astra on Fast | $17.60 | $88.00 | 10,560 | 9,607 | 9,85,248 |
| GPT-6.1 Terra on Fast | $2.20 | $11.00 | 1,320 | 1,201 | 1,23,156 |
| GPT-5.2 on Fast (ends 11 Jan) | $3.50 | $28.00 | 2,520 | 2,293 | 2,35,116 |
| GPT-5.2 on Standard | $1.75 | $14.00 | 1,260 | 1,146 | 1,17,558 |
| GPT-5.4 mini on Fast (ends 11 Jan) | $1.65 | $9.90 | 1,056 | 961 | 98,525 |
| GPT-5.4 mini on Standard | $0.825 | $4.95 | 528 | 480 | 49,262 |
Dollar rates are from OpenAI's pricing page, checked on 20 November 2026. Euro figures use the ECB reference rate for 19 November 2026 of US$1.0992 per euro; rupee figures use the ECB cross of ₹102.55 per euro divided by US$1.0992, which gives ₹93.30 per dollar after rounding to two decimals. Both are rounded to whole units. Applying the uplift to Fast rates is our reading of the pricing page, which applies it to regional endpoints without excluding any tier.
What the table shows:
- GPT-5.4 Fast users save either way. Moving to GPT-6.2 Sol on Fast cuts the example bill by 25%, from $3,520 to $2,640, if token counts stay the same. Dropping to GPT-5.4 Standard halves it.
- GPT-5.2 Fast users pay a little more for speed on Sol. Sol on Fast costs $2,640 against $2,520 today, about 5% more, because GPT-5.2 carries no uplift. Sol on Standard costs about 5% more than GPT-5.2 Standard.
- GPT-5.4 mini users have no cheap GPT-6.2 equivalent. Sol on Standard costs $1,320, 25% more than mini on Fast today and 2.5 times mini on Standard. GPT-6.1 Terra on Fast lands at the same $1,320.
- Token counts will not stay the same. The reasoning-effort change alone can raise output tokens, and the prefix-edit behaviour can raise both. Run a sample of real traffic and read the usage object before you trust any row.
For overnight or queue-based work there is a cheaper line again: Sol on an eight-hour background window costs 40% below Standard, which would bring the example to $792 a month with the uplift ($1,320 x 0.6). The background guide does not say whether discounted windows run on EU-residency projects, so confirm that with OpenAI first, and use it only for work nobody is waiting for.
A 52-day plan to 11 January
| Window | Action | Done when |
|---|---|---|
| 20 to 27 Nov | Inventory EU Fast traffic by dashboard, code search and response logs | Every affected call has an owner |
| 20 to 27 Nov | Set EU project default tier to Standard where calls do not set it | No call inherits Fast silently |
| 28 Nov to 11 Dec | Run GPT-6.2 Sol, Fast and Standard, against a golden set of real prompts | Quality, latency and tokens recorded |
| 28 Nov to 11 Dec | Fix reasoning effort, sampling, caching and tool-calling differences | Tests pass on GPT-6.2 |
| 12 to 18 Dec | Ship behind a flag, ramp gradually, watch the service_tier field | Production share rising without errors |
| 19 Dec to 4 Jan | Holiday freeze buffer; finish stragglers on Standard | No affected call left on Fast |
| 4 Jan | Internal deadline, one week before shutdown | Monitoring alerts on any EU priority tier |
Ramp carefully if you move traffic onto GPT-6.2 Fast. OpenAI says traffic that ramps too fast can be downgraded to Standard speed and Standard price, and suggests that once you reach 1 million input tokens per minute you grow by no more than 50% every 15 minutes, shifting traffic with feature flags over hours. Moving a whole region from one model to another in one deploy is exactly the pattern that guidance warns against.
Pick 4 January as your internal deadline. The window crosses the December holidays, when change freezes and thin staffing make a late fix harder. Anything that slips can still fall back to Standard on the same model, which is the one change you can make in minutes.
Which option suits which workload
| Workload | Current setup | Suggested route | Why |
|---|---|---|---|
| Customer-facing chat or copilot | GPT-5.4 on EU Fast | GPT-6.2 Sol on Fast | Keeps the low-latency tier and costs less per token |
| Agent with tool calls | GPT-5.4 or 5.2 on EU Fast | GPT-6.2 Sol on Fast via Responses | Tool calling on GPT-6.2 needs Responses |
| Internal summarisation or drafting | Any of the three on EU Fast | Same model on Standard | Latency rarely matters; bill halves |
| High-volume classification | GPT-5.4 mini on EU Fast | GPT-5.4 mini on Standard, then test | No GPT-6.2 model matches the mini price |
| Hard reasoning or coding | GPT-5.4 on EU Fast | Test GPT-6.2 Astra against Sol | Astra costs four times Sol; prove the gain |
| Overnight document jobs | Any on EU Fast | Batch, or a background window if EU-eligible | Fast was the wrong tier for this work |
The routes follow OpenAI's documented support and the rates above. Treat them as starting points for testing, not rulings.
Data protection angle
The notice changes speed and price, not where data is processed. Both named routes keep EU residency. That is the point to make to your data protection officer before anyone suggests a shortcut: moving latency-sensitive traffic off the EU endpoint to keep Fast on GPT-5.4 would change where customer content is processed and would need its own GDPR assessment.
Two scope notes from the data residency guide are worth repeating in that conversation. Residency covers customer content, not system data such as account details, usage statistics and billing records, which may be stored and processed outside the region. And the EU region covers the EEA plus Switzerland.
India-specific considerations
Indian software firms often run European clients' workloads on an EU-residency OpenAI project so that processing stays in the EEA. Those projects are in scope exactly as a European company's would be, and the inventory above applies.
For Indian end users the picture is different. OpenAI's India region at in.api.openai.com supports regional processing only for gpt-6-luna, on Standard pay-as-you-go, through a project created in that region, with no hosted tools. None of the three affected models or the GPT-6.2 family can process in-region in India, so the EU deadline does not change India-resident options. If your concern is the Digital Personal Data Protection Act rather than GDPR, our guide to DPDP-ready multi-region cloud architecture covers that design.
For budgeting in rupees, the example workload on GPT-5.4 EU Fast costs about ₹3.28 lakh a month at the 19 November ECB rate, against about ₹2.46 lakh on GPT-6.2 Sol Fast and ₹1.64 lakh on GPT-5.4 Standard.
How eCorpIT can help
eCorpIT is a Gurugram-based technology organisation, founded in 2021, assessed at CMMI Level 5 and MSME certified, with senior-led engineering teams working across AWS, Microsoft and Google platforms. Our AI and machine learning services team can inventory your EU-residency OpenAI traffic, test GPT-6.2 Sol and Astra against your own prompts, and ship the tier or model change behind a flag before 11 January. Talk to us at /contact-us/.
Last updated: 20 November 2026.