Vertex AI has been renamed Gemini Enterprise Agent Platform, and if you call Gemini through a Google Cloud project your 2.5 models stop tomorrow — 20 October 2026. On the Gemini Developer API the same three model IDs have no shutdown date at all. Three costs apply only to the Cloud side: any endpoint carrying a data-residency guarantee is a non-global endpoint and costs exactly 10% more per token; Gemini 4 Argon’s promotional rate arrives as 50% credits back on net spend, not a lower invoice; and from 13 November 2026 you are billed for long completions that fail. From tomorrow the platform’s only non-Flash Gemini is Argon, at $4/$20 from 1 December. Google Cloud’s own named replacement for 2.5 Pro is a Flash model.
Which Google do you buy from?
Everything below depends on this answer. The two paths share model IDs and marketing while running separate calendars, rate cards and terms of service.
| Signal | Gemini Enterprise Agent Platform (ex-Vertex AI) | Gemini Developer API |
|---|---|---|
| Endpoint | *-aiplatform.googleapis.com, or a regional/multi-region host | generativelanguage.googleapis.com |
| Auth | Cloud service account, ADC, workload identity | AI Studio API key |
| Billing | A Cloud billing account, on your GCP invoice | Google’s API billing, separate |
| Project | A GCP project with the Agent Platform API enabled | An AI Studio key, no GCP project needed |
| Residency, CMEK, VPC-SC, AXT | Available | Not available |
| 2.5 shutdown | 20 October 2026 | No date published |
| Terms | Google Cloud Platform ToS | Gemini API ToS — “not a Google Cloud offering” |
If you are not sure, check the invoice. Gemini usage appearing as a line item on a Google Cloud bill means the 20 October date is yours. Enterprise buyers are disproportionately on that side, because residency, CMEK and VPC-SC exist only there — which puts the shortest clock on the organisations with the longest change-control cycles.
The dagger footnote on every Agent Platform model page is the sentence that decides your exposure:
“Listed retirement dates refer to retirement of support in Gemini Enterprise Agent Platform. Models may remain accessible through the Gemini API after these dates have passed. The Gemini API is not a Google Cloud offering and is subject to its own terms of service.”
The Agent Platform retirement calendar
Verified 19 October 2026 against the model versions and lifecycle page, which Google itself stamps “Last updated 2026-10-19”.
| Model ID | Released | Agent Platform retirement | Google’s named replacement |
|---|---|---|---|
gemini-2.5-pro | 17 Jun 2025 | 20 Oct 2026 | gemini-3.8-flash or gemini-3.5-flash |
gemini-2.5-flash | 17 Jun 2025 | 20 Oct 2026 | gemini-3.8-flash or gemini-3.5-flash-lite |
gemini-2.5-flash-lite | 22 Jul 2025 | 20 Oct 2026 | gemini-3.8-flash, gemini-3.5-flash-lite or Gemma 4 |
gemini-3.6-flash | 21 Jul 2026 | 19 Nov 2026 | gemini-3.8-flash |
veo-3.1-generate-001 / -fast | 17 Nov 2025 | 17 Nov 2026 or later | — |
gemini-live-2.5-flash-native-audio | 12 Dec 2025 | 13 Dec 2026 | — |
gemini-3.7-flash | 13 Aug 2026 | 28 Jan 2027 | gemini-3.8-flash |
gemini-2.5-flash-image | 2 Oct 2025 | 15 Mar 2027 | gemini-3.1-flash-lite-image |
gemini-3.1-flash-lite | 7 May 2026 | 7 May 2027 or later | — |
gemini-3.5-flash | 19 May 2026 | 19 May 2027 or later | — |
gemini-3.5-flash-lite | 21 Jul 2026 | 21 Jul 2027 or later | — |
gemini-3.8-flash | 2 Sep 2026 | 2 Sep 2027 or later | — |
gemini-4-argon | 13 Oct 2026 | 13 Oct 2027 or later | — |
Two rows deserve more attention than the 2.5 ones, because nothing has flagged them. gemini-3.6-flash retires on 19 November 2026 — one month away, for a model released in July 2026. And gemini-3.7-flash retires on 28 January 2027, four weeks after every Flash rate doubles on 1 January. Neither carries a shutdown date on the Gemini Developer API, because on that side both were deprecated on 8 October and silently rerouted rather than given dates — Google’s deprecation page now lists each with a note that the endpoint is not shutting down and requests are routed onward. The 3.6 Flash deadline, the replacement arithmetic and the thinking floor it hides are worked through in Migrate off Gemini 3.6 Flash on Google Cloud.
All of these dates are in our model retirement tracker, which also exports an .ics calendar.
The models the two platforms disagree about
This is the part that breaks plans, and it is not a case of one platform lagging the other. In four cases the schedules are inverted: the model that is dead on one is healthy on the other.
| Model | Gemini Developer API | Agent Platform |
|---|---|---|
gemini-3.5-flash | Gone — deprecated 8 Oct 2026, rerouted to 3.6 Flash | Supported to 19 May 2027 or later |
gemini-3.7-flash | Gone — deprecated 8 Oct 2026, rerouted to 3.8 Flash | Supported, firm date 28 Jan 2027 |
gemini-3.6-flash | Active, no shutdown date — the reroute target | Retires 19 Nov 2026 |
gemini-3.1-flash-image | Shuts down 29 Oct 2026 | Supported to 28 May 2027 or later |
gemini-2.5-pro / -flash / -flash-lite | No date; new projects blocked since 18 Sep 2026 | Retires 20 Oct 2026 |
gemini-3.1-pro-preview | Available (preview) | Not listed at all |
veo-3.1 | -preview only, shuts down 22 Oct 2026 | GA -001, 17 Nov 2026 or later |
The practical reading: a team that standardised on 3.6 Flash because the Developer API routes 3.5 Flash traffic there has picked, on Cloud, the one Flash model with a date inside 30 days — while the model it avoided is supported for another eighteen months. And a Veo user told the Gemini API offers nothing GA has, on Cloud, a generally available veo-3.1-generate-001, which closes an open question on our Veo 3.1 migration guide.
The three costs that exist only on Google Cloud
1. Residency costs exactly 10%, and nothing labels it
The Agent Platform rate card splits every 3.x and 4-generation model into Global and Non-global rows, and the non-global row is always exactly 10% higher.
| Model | Global in / out | Non-global in / out |
|---|---|---|
| Gemini 4 Argon (to 30 Nov 2026) | $2.00 / $10.00 | $2.20 / $11.00 |
| Gemini 4 Argon (from 1 Dec 2026) | $4.00 / $20.00 | $4.40 / $22.00 |
| Gemini 3.8 Flash (to 31 Dec 2026) | $0.75 / $3.75 | $0.825 / $4.125 |
| Gemini 3.8 Flash (from 1 Jan 2027) | $1.50 / $7.50 | $1.65 / $8.25 |
| Gemini 3.5 Flash-Lite | $0.30 / $2.50 | $0.33 / $2.75 |
| Gemini 3.1 Flash-Lite | $0.25 / $1.50 | $0.275 / $1.65 |
| Gemini 2.5 Pro / Flash / Flash-Lite | single-priced — no split | single-priced — no split |
Per million tokens. Prices verified 19 October 2026 against the Agent Platform pricing page.
The surcharge is never described as a residency premium, but that is its effect. Google’s residency documentation states that “Global endpoints route and process data anywhere” and that data submitted to one “may be processed in any Google Cloud location” — so a residency commitment means a regional or multi-region endpoint, which is the non-global column.
The sting is in the last row. The 2.5 generation has no region split at all. A European team moving off 2.5 Flash-Lite in europe-west4 is not paying the generation increase; it is paying the generation increase and then 10%. On a reference month of 50,000 requests at 4,000 input / 600 output tokens — 200M input, 30M output:
| Model | Global | Non-global (EU) | vs. 2.5 Flash-Lite |
|---|---|---|---|
| 2.5 Flash-Lite | $32.00 | $32.00 | baseline |
| 3.1 Flash-Lite | $95.00 | $104.50 | 3.3x |
| 3.5 Flash-Lite (recommended) | $135.00 | $148.50 | 4.6x |
| 3.8 Flash (2026) | $262.50 | $288.75 | 9.0x |
| 3.8 Flash (2027) | $525.00 | $577.50 | 18.0x |
| 2.5 Pro | $550.00 | $550.00 | — |
| 4 Argon (to 30 Nov) | $700.00 | $770.00 | — |
| 4 Argon (from 1 Dec) | $1,400.00 | $1,540.00 | — |
The headline 4.2x that a global-endpoint user faces on the recommended Flash-Lite migration is 4.6x under EU residency. For a 2.5 Pro user there is no Pro-tier landing spot on Cloud at all: $550 becomes $770 now and $1,540 from December on Argon, or you drop to Flash. That arithmetic, and the thinking-token increase that sits on top of it, is worked through in Migrating off Gemini 2.5; the per-token rates are in our API cost calculator.
2. Argon’s promotional rate is credits, not a discount
The Developer API sells Gemini 4 Argon at $2/$10 until 30 November and $4/$20 after. The Agent Platform shows the same $2/$10, with an asterisk:
“Promotional pricing provided through 50% credits back on net spend on select models within a given period. Promotional pricing does not apply to Gemini 4 Argon Cyber Permissive.”
That is a different instrument. Your invoice is computed at the standard rate and half comes back as Google Cloud credits. The per-token headline matches, but credits are not cash: they are scoped to Google Cloud, they generally carry expiry terms, and they do not reduce the invoice your finance team reconciles. Model Argon’s launch window as deferred credit rather than a discount, and confirm the expiry before counting it. What Argon costs and when it doubles covers the Developer API side.
3. From 13 November, failed long completions are billable
The same pricing page carries two statements that contradict each other. Near the top:
“You’re charged only for requests that return a 200 response code. Requests returning any other response codes, such as 4xx and 5xx codes, aren’t charged for the input or output.”
And in the Gemini 4 footnotes:
“When a completion exceeds 32K tokens in length, users will be charged for retrievable token usage even if the request ultimately fails due to client or server timeouts. This will take effect on November 13 2026.”
So from 13 November 2026 the 200-only rule has an exception, and it lands exactly where bills are largest: Argon generates up to 1,048,576 output tokens, and long decodes are the requests most likely to hit a client or server timeout. A 500K-token completion that times out at 480K is currently free and will cost $9.60 at the global rate from 1 December. Anything that retries long generations should cap max_output_tokens, raise client timeouts above the generation’s realistic ceiling, and stop treating a timeout as a free retry. The same cliff on the subscription side is tracked in our pricing tracker.
Grounding: the free allowance is attached to the models retiring tomorrow
Google Search grounding is the one place where Cloud is dramatically cheaper than the Developer API — for now.
| Agent Platform | Gemini Developer API (3.x) | |
|---|---|---|
| Per 1,000 grounded prompts | $35 | $14 |
| Free allowance | 1,500/day (Flash and Flash-Lite, combined); 10,000/day (Pro) | 5,000/month, shared across all 3.x |
1,500 a day is roughly 45,000 a month. A team running 45,000 grounded requests pays $0 on the Agent Platform and about $560 a month on the Developer API — the opposite of the usual assumption that Cloud carries the premium.
The catch is which models the allowance names. Google’s wording is “Gemini 2.0 Flash, 2.5 Flash and 2.5 Flash-Lite include a combined 1,500 Grounding Prompts per day at no additional charge. Gemini 2.5 Pro includes 10,000 Grounding Prompts per day.” Every model in that sentence either retires tomorrow or is already gone. No 3.x or 4 model is listed as carrying a free grounding allowance on Cloud, and the page has not been updated to extend one. That may be an editing lag rather than a policy, but if your Cloud workload depends on free grounding, the documented basis for it expires on 20 October 2026. Price a 3.x grounded request at the full $35 per thousand until Google says otherwise.
What the rename means for your code
Three things, in descending order of urgency.
The SDK deadline has already passed. The Generative AI module in the Vertex AI SDK was deprecated on 24 June 2025 and, per Google’s own table, “will be removed on June 24, 2026.” That date is four months behind us. If you are still importing vertexai.generative_models, you are running on a removed module — migrate to the Google Gen AI SDK, which is the one path Google documents.
The docs are frozen and the old URLs redirect. Vertex AI pages now carry “Vertex AI documentation is no longer being updated,” and cloud.google.com/vertex-ai/generative-ai/docs/models/gemini/2-5-pro lands on docs.cloud.google.com/gemini-enterprise-agent-platform/models/gemini/2-5-pro. Any runbook or vendor-assessment questionnaire citing a Vertex URL now cites a redirect to a differently-named product, which matters when a compliance reviewer follows the link.
Nothing in your request path changes. Endpoints, project IDs, service accounts, quotas and billing are untouched. There is no migration for the rename itself and no date to plan around, because Google has published none. The work is the model calendar, not the brand.
Checklist
- Determine which Google each workload bills through. Check the invoice, not the code. A single organisation commonly has both, with different exposure.
- If any workload is on Cloud and calls
gemini-2.5-*, it breaks tomorrow. That is today’s work, not this quarter’s. - Diary 19 November for
gemini-3.6-flashand 28 January for3.7-flash. Both are Cloud-only dates that no Developer API tracker will surface. - Re-price every Cloud workload in a non-global region at +10%. If residency is a requirement, this is not optional and it is not in the headline rate.
- Reclassify Argon’s launch pricing as credits in any Cloud forecast, and confirm the credit expiry terms in writing.
- Cap
max_output_tokensand raise client timeouts before 13 November on anything generating long completions. - Stop assuming free grounding on 3.x. Budget $35 per thousand on Cloud until the allowance is documented for a model you can still call.
- Pin lifecycle checks to the Agent Platform page, not Google’s Developer API deprecation page, for anything billed through Cloud.
For the model-by-model replacement arithmetic see Migrating off Gemini 2.5, and for the Veo side Migrate off Veo 3.1. Teams buying under EU residency should read the AI APIs you can buy with EU data residency and, for Claude specifically, Bedrock vs the Agent Platform. Context-window pricing cliffs are in long-prompt API costs. Product-level reviews: Gemini and Gemini CLI.
What changed
- 20 Oct 2026: Corrected the 3.7 Flash line — its 28 January 2027 retirement falls four weeks after the 1 January rate doubling, not two days. Both 3.5 and 3.7 Flash now appear on Google’s Developer API deprecation page with reroute notes. Full 3.6 Flash guide: Migrate off Gemini 3.6 Flash on Google Cloud.
- 19 Oct 2026: Page created. The Agent Platform retirement table, the ±10% global/non-global split, the Argon credits footnote, the 13 November failed-completion billing change and the grounding allowance wording were each verified by direct fetch of Google’s pages on this date; the model-versions page carries Google’s own “Last updated 2026-10-19” stamp.
Frequently asked questions
Is Vertex AI being shut down?
No — it is being renamed, and the documentation has already moved. Google Cloud's Vertex AI pages now open with the banner 'Vertex AI documentation is no longer being updated' and 'Vertex AI's services are now part of Gemini Enterprise Agent Platform.' Requests to the old documentation URLs now redirect: cloud.google.com/vertex-ai/generative-ai/docs/models/gemini/2-5-pro resolves to docs.cloud.google.com/gemini-enterprise-agent-platform/models/gemini/2-5-pro. Google publishes no date for the transition anywhere we could find, and no deprecation notice for the Vertex AI name itself. What has a date is the SDK: the Generative AI module in the Vertex AI SDK was deprecated on 24 June 2025 and removed on 24 June 2026, so that deadline has already passed. Your API endpoints, project IDs and billing are unchanged; treat this as a documentation and branding migration with a model-retirement calendar attached, not a service shutdown.
Does the 20 October 2026 Gemini 2.5 retirement apply to my API key?
Only if you call Gemini through a Google Cloud project. The Agent Platform model pages carry the notice 'This model is being retired on October 20th, 2026' and a Retirement date field reading 'October 20, 2026' for gemini-2.5-pro, gemini-2.5-flash and gemini-2.5-flash-lite. A dagger footnote reconciles the two Googles: 'Listed retirement dates refer to retirement of support in Gemini Enterprise Agent Platform. Models may remain accessible through the Gemini API after these dates have passed. The Gemini API is not a Google Cloud offering and is subject to its own terms of service.' So if you authenticate with a Cloud service account, bill through Cloud, or rely on VPC-SC, CMEK or data residency, the date is yours. If you call generativelanguage.googleapis.com with an AI Studio key, it is not — there the 2.5 models have no shutdown date at all, and your problem is the access gate instead.
Why is EU data residency 10% more expensive on Google Cloud?
Because residency forces you off the global endpoint, and the Agent Platform rate card prices every non-global region exactly 10% above global for the 3.x and 4 generations. Gemini 4 Argon is $2.00 input / $10.00 output per million tokens on global and $2.20 / $11.00 non-global; Gemini 3.8 Flash is $0.75 / $3.75 against $0.825 / $4.125; Gemini 3.5 Flash-Lite is $0.30 / $2.50 against $0.33 / $2.75. Google's data-residency documentation is explicit that 'Global endpoints route and process data anywhere' and that data sent to one 'may be processed in any Google Cloud location', so any residency commitment means a regional or multi-region endpoint — which is the non-global column. The surcharge is not labelled as a residency premium anywhere, which is why it tends to be missed in budgets. Note that the 2.5 generation has no global/non-global split at all: it is single-priced, so migrating off 2.5 in an EU region costs the generation jump plus 10%, not just the generation jump.
What replaces Gemini 2.5 Pro on Google Cloud?
A Flash model, according to Google's own table. The Agent Platform model-versions page lists gemini-2.5-pro's replacement as 'gemini-3.8-flash or gemini-3.5-flash' — a tier down, not across. Gemini 4 Argon is on the platform and is not named as the replacement for anything. From 20 October 2026 the platform's entire Gemini lineup is Argon plus Flash and Flash-Lite variants: there is no Pro-tier Gemini on Google Cloud at all, and Gemini 3.1 Pro Preview, which exists on the Developer API, is not listed. So a 2.5 Pro user on Cloud chooses between dropping to Flash or paying for a frontier model. On a 200M-input / 30M-output month, 2.5 Pro cost $550 on global; Argon costs $700 through 30 November and $1,400 from 1 December, when its launch rate expires.
Do the Agent Platform and the Gemini Developer API agree on which models are alive?
No, and in several cases they are exactly inverted. Gemini 3.5 Flash was deprecated on the Developer API on 8 October 2026 and silently rerouted to 3.6 Flash; on the Agent Platform it is a supported model with a retirement date of 'May 19, 2027 or later'. Gemini 3.7 Flash was likewise rerouted away on the Developer API; on Cloud it is live with a firm date of 28 January 2027. Gemini 3.6 Flash — the model the Developer API now routes 3.5 Flash traffic to — retires on Cloud on 19 November 2026, a month from now. And gemini-3.1-flash-image shuts down on the Developer API on 29 October 2026 while the Agent Platform lists it to 'May 28, 2027 or later'. Treat the two platforms as two vendors that happen to share model IDs, and read the calendar for the one you actually bill through.
Does Google Cloud's deprecation policy cover Gemini model retirements?
Not on the page that defines it. The Generative AI deprecations page cites the 'Google Cloud Platform Terms of Service (section Discontinuation of Services)' and states that 'the deprecation policy only applies to the services, features, or products listed therein.' The only row in that table is the Generative AI module in the Vertex AI SDK. No Gemini model appears, including the three retiring on 20 October. What the model-versions page offers instead is a product commitment rather than a contractual one: models in the main table are 'available for at least 12 months after initial release', retirement timelines 'won't be moved to an earlier date than what is listed', and short-term models get 'at least 45 days to migrate' once a date is posted. Those terms are better than the Gemini Developer API's — which publishes no minimum notice and calls its dates 'earliest possible' — but they live in documentation, not in the deprecation policy.