Gemini 4 Release Date: Confirmed, Not Released

The Gemini 4 release date has not been announced. Gemini 4 is real as a Google-named training program, but it is not a released model. On 2026-07-21, Google said that it had started pre-training Gemini 4. As of 2026-09-20, Google's Gemini API model catalog, pricing page, and changelog contain no public gemini-4, gemini-4-pro, or gemini-4-flash product; the Models API is the account-level check when that changes.

That distinction is the whole story. Google has confirmed the work, not a release date, model family, specification, price, or callable API ID. A roadmap fact is not yet a product contract.

This guide records the narrow official statement, separates it from the release claims circulating around it, shows how to check your own model list, and gives you a current route instead of asking you to wait for an unannounced one.

Executive answer

What Google actually confirmed

Google named Gemini 4 once in its 2026-07-21 announcement for Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber. The relevant sentence is short:

“We have started our most ambitious pre-training run yet, for Gemini 4, and are excited by the progress.”

That sentence establishes three things:

1. Google uses the name Gemini 4.
2. A pre-training run had started by 2026-07-21.
3. Google described it as its most ambitious pre-training run at that time.

It does not establish a completion date, launch date, model ID, tier names, context window, modalities, benchmark result, price, region, quota, consumer plan, or API availability. Every one of those needs another source.

The surrounding announcement makes the boundary even clearer. Google described Gemini 3.6 Flash and 3.5 Flash-Lite as products developers could use, and said Gemini 3.5 Pro was testing with partners. Gemini 4 appeared only in the forward-looking sentence about the next generation.

Why pre-training is not a release

Three different states are being collapsed in many Gemini 4 headlines:

Gemini 4 is at the first state in the public record. Treating that as the third creates broken configuration, invented budgets, and comparisons against a model nobody can reproduce.

The distinction also protects against a second source of confusion: the Gemini consumer app and the Gemini API are separate release surfaces. A name in an app picker, experiment, arena, or partner preview would not by itself define an API model ID.

Is there a Gemini 4 release date?

No. Google's announcement contains no date or launch window, and the API changelog through 2026-09-20 has no Gemini 4 entry.

Late-2026 and October predictions appear in search results, prediction-market discussions, and secondary articles. They are forecasts, not Google commitments. Even a forecast that eventually lands on the correct month does not become evidence retroactively; developers need a model ID, rate card, and supported contract before a date is operationally useful.

Watch these five surfaces instead of a rumor calendar:

1. Google's Gemini API model catalog.
2. The official Gemini API changelog.
3. The Gemini API pricing table.
4. The Models API response for your own account and region.
5. A model page or migration guide that documents the exact ID and parameters.

Until at least the first four agree, do not put a Gemini 4 identifier into production configuration.

Gemini 4 Pro, Flash, pricing, and specs are not announced

Autocomplete demand makes the names sound established. The official record does not.

Google has used Pro, Flash, and Flash-Lite elsewhere. That naming history makes the guesses understandable, but a naming pattern is not a contract. The final family could use those names, different names, or a staged release that exposes only one variant first.

The same rule applies to inherited specifications. Do not copy Gemini 3.8 Flash's 1,048,576-token input limit, 65,536-token output limit, thinking levels, or rates into a Gemini 4 spreadsheet. They describe a current model, not an unreleased generation.

Rumor log: what is evidence and what is not

The important pattern is independence. Ten articles that cite the same anonymous claim are one source, not ten confirmations. A screenshot without a request, response, account scope, and exact endpoint is a lead to investigate, not a specification to publish.

Check Gemini 4 availability yourself

Google's Models API lists the models available to your Gemini API credential. Keep the key in an environment variable and send it in the documented header rather than placing it in a URL:

set -o pipefail
curl --fail-with-body --silent --show-error \
  https://generativelanguage.googleapis.com/v1beta/models \
  -H "x-goog-api-key: $GEMINI_API_KEY" \
| jq -r '.models[]?.name | sub("^models/"; "") | select(startswith("gemini-4"))'

An empty result means that credential did not receive a matching ID at that moment. It does not rule out a private preview elsewhere, but it does rule out configuring that account as though the model were available.

Check the gateway you actually plan to call as a separate step:

set -o pipefail
curl --fail-with-body --silent --show-error https://api.omniakey.com/v1/models \
  -H "Authorization: Bearer $OMNIAKEY_API_KEY" \
| jq -r '.data[]?.id | select(startswith("gemini-4"))'

Provider and gateway lists answer different questions. Google's response shows what Google exposes to that credential; OmniaKey's response shows what OmniaKey currently routes. A model can appear on one surface before another.

Use exact matching. A response mentioning “Gemini 4” in descriptive text, or an ID such as gemma-4-*, is not a gemini-4-* route.

What to use while Gemini 4 is training

For general text, coding, and tool-using agent work, the current documented route is gemini-3.8-flash.

Those are Gemini 3.8 facts, not Gemini 4 predictions. The Gemini 3.8 Flash pricing guide separates Google Standard, Batch, Flex, Priority, and the independent OmniaKey quote, then calculates retry-aware cost per accepted task. The model page owns the current gateway rate and route status.

If your requirement is simply “the next strong Google model,” waiting is not an evaluation plan. Run the workload against the current model with a fixed acceptance gate. When Gemini 4 becomes callable, replay the same set and compare accepted-task cost, latency, tool reliability, and failure categories instead of comparing launch adjectives.

Prepare for Gemini 4 without pinning a fictional ID

You can make a future migration cheap without pretending the model already exists:

1. Keep the model ID in runtime configuration, but keep its current value valid.
2. Add an exact model-list preflight to deployment smoke tests.
3. Freeze a representative evaluation set and the commands or rubric that decide acceptance.
4. Record input, cached input, thinking, visible output, retries, tools, latency, and final cost.
5. Isolate Gemini-native fields from OpenAI-compatible fields so a new contract does not leak across clients.
6. Require an explicit rollout decision after the model page and rate card exist; do not auto-switch on a name sighting.

This turns launch day into a measured routing decision. Without the baseline, even a real Gemini 4 benchmark tells you little about your prompts, tools, permissions, and retry policy.

What will count as a Gemini 4 release

We will update this page from status guide to release guide when the public evidence supports it. The minimum useful release record is:

1. A Google model page naming an exact Gemini 4 ID.
2. A dated changelog or launch announcement.
3. A pricing row or an explicit free-preview boundary.
4. A Models API response that exposes the ID to an eligible account.
5. Documented limits, supported methods, regions, and lifecycle state.

OmniaKey availability is a sixth, separate check. Google releasing a model does not mean every gateway routes it immediately. This article will not claim an OmniaKey route until the live catalog and request path support it.

Common mistakes

Treating an official name as an official product

Gemini 4 is officially named. That validates the generation, not a public SKU. Preserve both halves of the sentence.

Turning family history into model IDs

gemini-4-pro and gemini-4-flash look reasonable because older families use those tiers. Sending either string today is still a guess.

Mixing the Gemini app with the Gemini API

Consumer plan access, an experimental picker, Google AI Studio, Vertex AI, and the Gemini API can move on different schedules. Always name the surface behind an availability claim.

Publishing benchmark comparisons before access exists

There is no reproducible public Gemini 4 run to compare with GPT, Claude, or Gemini 3.8. Leaked scores without prompts, settings, model identity, and artifacts are not a review.

Budgeting with inherited prices

Gemini 3.8 prices are a current baseline, not a forecast. Gemini 4 could introduce different tiers, consumption modes, long-context rules, or preview terms.

Frequently asked questions

Is Gemini 4 real?

Yes, as a Google-confirmed pre-training program. Google named it in an official post on 2026-07-21. It is not yet a released public model.

Has Gemini 4 been released?

No. As of 2026-09-20, Google has no Gemini 4 model page, pricing row, changelog release, or public API ID.

When is the Gemini 4 release date?

Google has not announced one. October, November, and “late 2026” dates are forecasts or rumors, not commitments.

Is Gemini 4 Pro confirmed?

No. Google confirmed the Gemini 4 generation but has not announced a Pro variant or gemini-4-pro model ID.

Is Gemini 4 Flash confirmed?

No. Flash is an established Gemini tier name, but Google has not announced gemini-4-flash.

How much will the Gemini 4 API cost?

No official price exists. Do not reuse Gemini 3.8 rates or consumer subscription prices as a forecast.

What is the Gemini 4 API model ID?

Google has not published one. gemini-4, gemini-4-pro, and gemini-4-flash are not current public IDs.

Which Gemini model should I use now?

Start with gemini-3.8-flash for current general text, coding, and agent workloads, then measure it against your own acceptance gate. Use the live catalog for route availability and the pricing guide for current cost modes.

Does OmniaKey support Gemini 4?

No public Gemini 4 route exists to support. OmniaKey currently lists gemini-3.8-flash; Gemini 4 will be evaluated only after Google publishes a callable contract.

Official sources and rumor audit

Official sources, checked on 2026-09-20:

• Google: Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber announcement — the primary Gemini 4 pre-training statement
• Google: Gemini API model catalog
• Google: Gemini API changelog
• Google: Gemini API pricing
• Google: Models API reference
• Google: Gemini 3.8 Flash model page
• OmniaKey live model catalog

Representative secondary coverage was reviewed only to trace the rumor surface: Fello AI's status overview and MindStudio's release-rumor article. Their forecasts and unattributed technical claims are not used as product facts here.

Evidence was checked on 2026-09-20. Release state, model IDs, prices, and limits can change; re-run the model-list checks and read the official rate card before changing production configuration.