If you call Gemini through a Google Cloud project, gemini-3.6-flash stops on 19 November 2026 — 30 days from now, for a model released in July 2026. On the Gemini Developer API the same model ID has no shutdown date at all, and is the model Google routes deprecated 3.5 Flash traffic into. The move is 3.8 Flash, and the rate card makes it free: 3.6 and 3.8 Flash are priced identically at every tier, so there is no generational increase to absorb. Two things are not free — the minimal thinking level disappears on 3.8 Flash, and 3.5 Flash is a trap: eighteen months of runway at 2.17x the cost, with a 20% output premium that survives January. Cheaper landing spot if your workload tolerates it: 3.5 Flash-Lite at 49% less.
First, work out whether this deadline is yours
This retirement applies to one of Google’s two Gemini platforms, and the two disagree about this exact model.
Google Cloud — retires 19 November 2026. The Gemini Enterprise Agent Platform model-versions table lists gemini-3.6-flash with a release date of 21 July 2026, a retirement date of 19 November 2026 and gemini-3.8-flash as the named replacement. Unusually, it sits in the platform’s short-availability group rather than the “at least 12 months after release” group that holds 3.8 Flash, 3.5 Flash and 3.5 Flash-Lite — so the model got roughly four months of supported life. The model’s own page repeats it in prose, telling readers that 3.6 Flash will be retired on 19 November 2026 and to upgrade to 3.8 Flash to avoid service disruptions.
Gemini Developer API — no date. Google’s deprecation page lists gemini-3.6-flash with its 21 July 2026 release date and nothing in the shutdown column. That page’s tables carry only four columns — model, release date, shutdown date, recommended replacement — and no deprecation dates at all; it also states that the shutdown dates it does publish are the earliest possible dates on which a model might be retired, with exact dates communicated later.
You are on the 19 November side if your calls are authenticated by a Cloud service account, go through the Vertex AI SDK, or sit behind VPC-SC, CMEK or a data-residency commitment. You are on the no-date side if you call generativelanguage.googleapis.com with an AI Studio key. Enterprise buyers are disproportionately in the first group, which is also the group with the longest change-control cycle. The full Cloud retirement calendar, the Argon credits footnote and the 13 November failed-completion billing change are in Vertex AI is now Gemini Enterprise Agent Platform.
Why this one is worse than a normal retirement
Because Google’s own Developer API guidance points at it.
gemini-3.5-flash and gemini-3.7-flash were deprecated on the Developer API on 8 October 2026 and silently rerouted rather than given shutdown dates. Google’s deprecation page now carries both, each with a note that the model is not selectable in Google AI Studio, the endpoint is not shutting down, and API requests are automatically routed onward — 3.5 Flash to gemini-3.6-flash, 3.7 Flash to gemini-3.8-flash.
Follow that guidance and the outcome is perverse. A team on 3.5 Flash is told, by Google, that its traffic now lands on 3.6 Flash. On Cloud, 3.6 Flash is the single Flash version with a date inside 30 days — while 3.5 Flash, the model it was moved off, is supported there until 19 May 2027 or later. The platform that gave you the instruction and the platform that enforces the deadline are, for billing purposes, two vendors sharing model IDs.
The second-order problem is that nothing will tell you. A Developer API deprecation scanner sees gemini-3.6-flash with an empty shutdown cell and reports no exposure. The date lives only in the Agent Platform’s model-versions table, which is a different site, a different product and a different page from the one most dependency tooling watches. Both of these dates are in our model retirement tracker, which also exports an .ics calendar.
The replacement map, priced
All figures per million tokens, global endpoint, standard tier. Prices verified 20 October 2026 against the Agent Platform pricing page.
| Model | Input | Output | Cached input | Cloud retirement |
|---|---|---|---|---|
gemini-3.6-flash (to 31 Dec) | $0.75 | $3.75 | $0.075 | 19 Nov 2026 |
gemini-3.8-flash (to 31 Dec) | $0.75 | $3.75 | $0.075 | 2 Sep 2027 or later |
gemini-3.8-flash (from 1 Jan) | $1.50 | $7.50 | $0.15 | — |
gemini-3.7-flash (to 31 Dec) | $0.75 | $3.75 | $0.075 | 28 Jan 2027 |
gemini-3.5-flash | $1.50 | $9.00 | $0.15 | 19 May 2027 or later |
gemini-3.5-flash-lite | $0.30 | $2.50 | $0.03 | 21 Jul 2027 or later |
gemini-3.8-flash-cyber | $1.50 | $7.50 | $0.15 | — |
Non-global endpoints are exactly +10% on every row, which is the figure that applies to any endpoint carrying a residency guarantee — the full EU picture, vendor by vendor, is in the AI APIs you can buy with EU data residency. Google’s footnote dates that split: non-global pricing took effect for generally available Gemini 3 and later families on 1 July 2026, and before that date global pricing applied to non-global endpoints.
Three things in that table are worth saying out loud.
3.6 → 3.8 Flash is free. Identical input, output, cached input and cache storage; identical introductory window; identical doubling on 1 January 2027; identical +10% non-global surcharge. Google’s header note confirms it in one sentence, bracketing 3.8, 3.7 and 3.6 Flash together under one introductory rate of $0.75 / $3.75 through 31 December 2026 and one standard rate of $1.50 / $7.50 from 1 January 2027. Generational migrations usually cost money — moving off Gemini 2.5 runs 4.2x to 9x — so a free one deserves to be taken while it is on offer.
3.5 Flash is the trap. It has the longest runway of any current Flash model and it is the most expensive, at $9.00 output against $3.75. On the reference month below that is $570.00 against $262.50. And the premium does not wash out in January: 3.5 Flash carries no dated increase, so when 3.8 Flash doubles, inputs match at $1.50 and 3.5 Flash’s output is still 20% higher. Buying lifecycle safety here costs 2.17x now and roughly 9% forever.
Long context is not a Flash problem. The Gemini 3 table splits every price into <= 200K input tokens and > 200K input tokens columns, and Google footnotes that a query over 200K tokens is charged at long-context rates on all tokens. For every Flash model the two columns are identical. The split only bites on Gemini 3.1 Pro Preview, where input goes $2.00 → $4.00 and output $12.00 → $18.00. If you are moving a long-prompt workload, the effective cost of long prompts is about caching, not about tiering.
What the move actually costs
Reference month: 50,000 requests at 4,000 input and 600 output tokens each — 200M input, 30M output.
| Model | Global | Non-global (EU) | vs. 3.8 Flash |
|---|---|---|---|
3.6-flash (today, retiring) | $262.50 | $288.75 | baseline |
3.8-flash (to 31 Dec) | $262.50 | $288.75 | — |
3.8-flash (from 1 Jan 2027) | $525.00 | $577.50 | 2.00x |
3.7-flash (to 31 Dec, bridge) | $262.50 | $288.75 | — |
3.5-flash (long runway) | $570.00 | $627.00 | 2.17x |
3.5-flash-lite (tier down) | $135.00 | $148.50 | 0.51x |
The shape of the decision: the deadline move costs nothing, the safe-looking move costs 2.17x, and the cheap move costs capability. That is an unusually clean set of options for a 30-day migration, and it argues for doing the free one now and treating the Flash-Lite question as a separate December exercise priced against the January doubling. Price your own shape in the API cost calculator; dated plan and rate changes are logged in the pricing tracker.
The thinking floor you lose
This is the one real cost inside a free migration, and it will not appear in any rate-card comparison.
Both platforms bill reasoning tokens as output. The Agent Platform labels the row Text output (response and reasoning); the Developer API labels it Output price (including thinking tokens). Reasoning is therefore the most expensive thing a Flash model does, and the setting that governs it changes between these two versions.
Google’s thinking documentation gives gemini-3.6-flash four levels — minimal, low, medium and high — and gemini-3.8-flash only three: low, medium and high. Both default to medium. There is no off value on either; for Gemini 3 models these are reasoning effort settings rather than a toggle.
So the migration has a hidden floor. A team that set thinking_level to minimal on 3.6 Flash specifically to hold down output spend cannot carry that configuration across. Its cheapest available setting on 3.8 Flash is low, and the difference lands entirely on the expensive half of the rate card.
Size it before you commit. On the reference month, each additional 200 reasoning tokens per request is 10M extra output tokens:
| Extra reasoning tokens / request | Added cost (to 31 Dec) | Added cost (from 1 Jan) |
|---|---|---|
| +200 | $37.50 | $75.00 |
| +500 | $93.75 | $187.50 |
| +1,000 | $187.50 | $375.00 |
Add 10% to each figure on a non-global endpoint. A workload that was deliberately pinned to minimal can therefore see its bill rise 15–70% on an unchanged rate card — which is exactly the failure mode that makes “the price did not change” a misleading thing to tell a finance team. The same mechanism, at a larger magnitude, is what makes the 2.5 generation’s migration more expensive than its rate card implies.
One further wrinkle from Google’s own documentation: the 3.8 Flash model page notes the model can use more tokens at higher effort levels, and offers lowering the effort level or switching to 3.7 Flash as mitigations. Both are priced identically to 3.6 Flash, so 3.7 Flash is a genuine same-cost bridge to 28 January 2027 — 70 days past the 3.6 Flash deadline, with its own price doubling 27 days before it retires.
Migration checklist
- Determine your platform first. Cloud service account, Vertex AI SDK, VPC-SC, CMEK or residency means 19 November 2026. An AI Studio key on
generativelanguage.googleapis.commeans no published date, and this page is a watch item rather than a deadline. - Grep for the string, not the family.
gemini-3.6-flashmay be a config value, an environment variable, a Terraform default, a notebook cell or a hardcoded fallback. Also search forgemini-3.5-flash, because on Cloud that is a live model and on the Developer API it is an alias pointing at the model that is about to die. - Change the model string to
gemini-3.8-flashand re-run evals. The rate card is identical, so there is no cost case to build — the only question is quality and token consumption. - Measure reasoning tokens at
lowagainst your 3.6 Flash baseline. If you were running minimal, assume an increase and size it with the table above before the change reaches production. - If 3.8 Flash burns materially more tokens, bridge on 3.7 Flash to 28 January 2027 at the same rate, and put that date in the calendar at the same time.
- Re-price everything in a non-global region at +10%, and re-price the whole workload at the 1 January 2027 standard rate rather than the introductory one. A budget built this quarter on $0.75 / $3.75 is a budget for ten weeks.
- Evaluate 3.5 Flash-Lite in December, when the doubling makes the capability trade worth 74% rather than 49%.
- Add a canary that calls every model you depend on from a freshly provisioned project, monthly. Access changes and reroutes do not show up in dashboards watching healthy production traffic.
For the lineup, plan tiers and what each Gemini subscription actually includes, see the Gemini review and Gemini CLI.
What changed
- 20 Oct 2026: Page created. The 19 November 2026 Cloud retirement, the 3.6/3.7/3.8 Flash price identity, 3.5 Flash’s $1.50 / $9.00 rate, the +10% non-global split and the 200K long-context columns were verified by direct fetch of Google’s Agent Platform pricing page and the Developer API deprecation page on this date; the deprecation page carries Google’s own “Last updated 2026-10-20” stamp and the model-versions page “Last updated 2026-10-19”. The thinking-level difference comes from Google’s thinking documentation and the 3.8 Flash model page.
Frequently asked questions
When exactly does Gemini 3.6 Flash stop working?
19 November 2026, but only if you bill through Google Cloud. The Gemini Enterprise Agent Platform — the renamed Vertex AI — lists gemini-3.6-flash with a release date of 21 July 2026 and a retirement date of 19 November 2026, naming gemini-3.8-flash as the replacement, and the model's own page carries a retirement notice telling you to upgrade to 3.8 Flash to avoid service disruptions. On the Gemini Developer API the same model ID has no shutdown date at all: Google's deprecation page lists it with its July release date and nothing in the shutdown column. So whether you have 30 days or indefinite time depends entirely on which Google your invoice comes from. If you authenticate with a Cloud service account, use the Vertex AI SDK, or have VPC-SC, CMEK or data residency configured, you are on the 19 November side.
What should I move to?
Gemini 3.8 Flash, in almost every case, and the reason is that it is free to do. On the Agent Platform rate card, 3.6 Flash and 3.8 Flash are priced identically at every tier and in every region column: $0.75 input and $3.75 output per million tokens on the global endpoint through 31 December 2026, $1.50 and $7.50 from 1 January 2027, with cached input at $0.075 then $0.15 and a flat +10% on non-global endpoints. There is no generational price increase to absorb, which makes this an unusually cheap migration — change the model string, re-run your evals, done. The two things that are not free are the thinking floor and the token appetite, both covered below.
Gemini 3.5 Flash is supported until May 2027 on Cloud. Should I move there instead for the longer runway?
No, and the price is why. The Agent Platform lists gemini-3.5-flash at 'May 19, 2027 or later' — roughly eighteen months of runway against 3.8 Flash's 'September 2, 2027 or later' — but it is priced at $1.50 input and $9.00 output per million on the global endpoint with no introductory discount and no dated change. That is 2.17x the cost of 3.8 Flash on a 200M-input / 30M-output month ($570.00 against $262.50) for as long as the promotional window runs. Worse, it does not converge. When every 3.x Flash rate doubles on 1 January 2027, 3.5 Flash's input matches 3.8 Flash at $1.50 but its output stays 20% higher at $9.00 against $7.50 — so the premium is permanent, not a timing artifact. You would be paying a durable surcharge to buy nine extra months on a model whose retirement date is a floor rather than a commitment anyway.
Why did my output bill rise after moving to 3.8 Flash when the rate card is identical?
Almost certainly the thinking floor. Both platforms bill reasoning as output — the Agent Platform's rate card labels the row 'Text output (response and reasoning)' — and Google's thinking documentation gives 3.6 Flash four levels (minimal, low, medium, high) while 3.8 Flash supports only low, medium and high. Both default to medium. So a team that explicitly set thinking_level to minimal on 3.6 Flash to hold the output bill down has no equivalent setting on 3.8 Flash; the floor rises to low, and there is no off switch on either model. The arithmetic is linear and easy to size: on 50,000 requests a month, every extra 200 reasoning tokens per request is 10M additional output tokens, which is $37.50 now and $75.00 from January on the global endpoint, $41.25 and $82.50 non-global. Measure your actual reasoning-token count on a sample at low before you size the budget rather than assuming the identical rate card means an identical invoice.
Google's docs suggest falling back to 3.7 Flash. Is that a real option?
As a bridge, yes; as a destination, no. Gemini 3.8 Flash's model page notes that it can consume more tokens at higher effort settings and offers two mitigations — lower the effort level, or switch to 3.7 Flash. On Cloud, 3.7 Flash is priced identically to 3.6 and 3.8 Flash and is supported until 28 January 2027, which is 70 days past the 3.6 Flash cutoff. So if your evals show 3.8 Flash burning materially more reasoning tokens than 3.6 Flash did, 3.7 Flash buys you a quarter at the same rate while you tune. Note what you are accepting: Google's own recommended workaround for its newest Flash model is a model that retires ten weeks later, and whose price doubles on 1 January, 27 days before it goes. Budget it as a bridge with a calendar entry, not as a second migration you get to skip.
Is there a cheaper landing spot than 3.8 Flash?
Yes, one tier down: Gemini 3.5 Flash-Lite at $0.30 input and $2.50 output per million on the global endpoint, with no introductory discount and no scheduled increase, supported to 'July 21, 2027 or later'. On the reference month that is $135.00 against 3.8 Flash's $262.50 — 49% less today, and 74% less once 3.8 Flash doubles in January, at which point Flash-Lite becomes the only Flash-family model on the Agent Platform whose price has not moved. The trade is capability, and it is a real one: Flash-Lite is a smaller model and the 3.6 Flash workloads most likely to survive the drop are classification, extraction and routing rather than long-horizon agent work. The honest sequencing is to move to 3.8 Flash before 19 November because it is free and the deadline is fixed, then evaluate Flash-Lite against your evals during December, when the January doubling gives the question a price tag.