Quick answer. Gemini 3.5 Pro and Flash are not officially announced or released as of May 19, 2026. Google's Gemini hub lists only Gemini 3, Gemini 3.1 Flash TTS, Gemini 3.1 Pro, and Embedding 2 — no “3.5.” Google I/O 2026 is live today, but no Gemini 3.5 name is confirmed. The current shipping model is Gemini 3 (Pro tier). Run that now.
If you searched for “Gemini 3.5 Pro release date” or “Gemini 3.5 Flash benchmarks,” here is the honest status: there is no official Gemini 3.5. No model card, no API model ID, no blog.google announcement, no benchmark page. “Gemini 3.5 Pro” and the “Cappuccino” codename are a pre-keynote leak, not a launch. Google I/O 2026 is happening today, which is exactly why speculation is loud right now — but loud speculation is not a release.
This page is the accurate status while everyone else extrapolates. It is written to refresh in place the moment Google officially announces something. Until then, treat every “Gemini 3.5” spec, benchmark, or price you see elsewhere as unverified.
Is Gemini 3.5 released yet?
No. As of May 19, 2026, Gemini 3.5 Pro and Gemini 3.5 Flash do not exist as released or officially announced products. Concretely, all of the following are absent:
- No
gemini-3.5-proorgemini-3.5-flashmodel ID in the Gemini API or Vertex AI model list. - No “Gemini 3.5” post on blog.google/products/gemini — the official Gemini hub lists Gemini 3, Gemini 3.1 Flash TTS, Gemini 3.1 Pro, and Embedding 2.
- No official model card, system card, or benchmark table from Google or DeepMind.
- No “3.5” tier inside the Gemini app's model picker.
When a name has zero footprint across the vendor's own API, docs, blog, and app, it is not a release. It is anticipation.
What Gemini models are actually shipping right now?
This is the part that matters if you have work to ship today. The current, real, supported Gemini lineup looks like this:
| Model | Tier | Best for |
|---|---|---|
| Gemini 3 (Pro) | Flagship reasoning | Hard reasoning, long-context, agentic and coding work |
| Gemini 3.1 Pro | Flagship, iterated | Deep Research / autonomous research-agent workflows |
| Gemini 3.1 Flash TTS | Speech | Expressive text-to-speech output |
| Gemini Embedding 2 | Embeddings (GA) | Retrieval, semantic search, RAG pipelines |
If you need a Gemini model in production today, Gemini 3 on the Pro tier is the answer for general reasoning and coding, and Embedding 2 is the answer for retrieval. Everything else labelled “3.5” is not something you can call.
Where did the Gemini 3.5 rumor come from?
The “Gemini 3.5 Pro” story has a few converging threads. None of them is an official source. Tracked honestly:
- A codename leak. Around May 15–17, 2026, secondary outlets (36Kr / Xinzhiyuan and follow-on coverage) reported a next-generation Gemini checkpoint internally codenamed Cappuccino, possibly surfacing on community benchmark platforms. This is a leak about an internal candidate build — not a product page.
- Release-cadence math. Google shipped Gemini 3, then Gemini 3.1 Pro and Gemini 3.1 Flash TTS. Observers extrapolate a “3.5” the way they did a “3.1.” Pattern-matching a version number is a guess, not a roadmap.
- I/O 2026 timing. Google I/O 2026's keynote is today, May 19, ~10am PT, led by Sundar Pichai. Pre-event press widely expects “major Gemini upgrades.” “Expected at a keynote” is the single most common way a rumor gets dressed up as a fact.
Treat all three as anticipated, not announced. A codename in a leak, a version-number pattern, and a keynote on the calendar do not add up to a shipped model.
What is “Cappuccino,” then?
Best available reading: “Cappuccino” is an internal codename for a candidate Gemini checkpoint, reported by secondary outlets ahead of I/O 2026. It is not confirmed to be “Gemini 3.5,” not confirmed to ship under any public name, and not confirmed to ship at all on any particular date. Internal checkpoints get renamed, merged, delayed, or dropped routinely.
Specifically, here is what is not established about Cappuccino, and which you should ignore if a post states it as fact:
- Any parameter count, context-window size, or architecture detail.
- Any benchmark score (coding, reasoning, multimodal, or otherwise).
- Any pricing, rate limits, or availability tier.
- Any firm release date — including “at I/O 2026.”
If you see a confident “Gemini 3.5 Pro benchmarks” table circulating, ask where the numbers came from. As of this writing there is no official source for any of them.
What might Gemini 3.5 bring if the pattern holds?
This section is explicit extrapolation from Google's public Gemini 3 → 3.1 trajectory. No numbers, because there are no verified numbers. Read it as “what a point release in this family has historically tended to touch,” not as a spec sheet:
- Tiered Pro / Flash split. Google has consistently shipped a heavier reasoning tier alongside a faster, cheaper tier (the Pro/Flash pattern). A future release would likely keep that split.
- Iterated coding and agentic behavior. The Gemini 3 line has been positioned around reasoning, long-context, and agent workflows (e.g., Deep Research on the 3.1 Pro side). A point release in the same family would plausibly continue that emphasis.
- Multimodal continuity. The family already spans text, speech (3.1 Flash TTS), and embeddings (Embedding 2). A successor would likely extend rather than abandon that surface.
That is the entire honest extrapolation. Anyone telling you the context window, the price per million tokens, or the SWE-bench number for “Gemini 3.5” before Google publishes a model card is guessing.
Companion guide
For how the broader open and frontier model field is moving — capabilities, variants, and how to choose — see our Gemma 4 complete guide for 2026.
Should you wait for Gemini 3.5 or use Gemini 3 now?
Use Gemini 3 now. There is no shipped “3.5” to wait for, and there is no published release date — so “waiting” is waiting for an unknown duration on an unconfirmed product. Practically:
- Building today? Target Gemini 3 (Pro) for reasoning and coding, Embedding 2 for retrieval. These are stable, documented, and callable.
- Architecting for the future? Keep your model layer swappable. Abstract the model ID behind config so that if and when Google ships a successor, switching is a one-line change — not a rewrite. This is good practice regardless of any specific release.
- Comparing vendors? Evaluate against what is shipping across the field today, not against a rumored model. For the wider landscape and how the major model families stack up, see our open-source LLM landscape for 2026.
The cost of waiting is concrete: a feature you could have shipped on Gemini 3 this week instead sits on a backlog pegged to an unconfirmed model with no date. The cost of building now is near zero, because a swappable model layer means you can adopt a successor in a config change the day it has a real model ID. Bias toward shipping on what exists.
Did Google announce Gemini 3.5 at I/O 2026?
As of this writing — the morning of May 19, 2026, with the I/O keynote scheduled but not yet delivered — Google has not confirmed a “Gemini 3.5” name or release. Pre-event press expects “major Gemini upgrades,” but expectation is not announcement, and a working name in leaks is not the name Google ships under.
This page updates in place once Google publishes something official. If you are reading this after the keynote and it still says “not confirmed,” that means there was no official model card or blog.google post at the time of the last update — not that we missed it.
How will you know when Gemini 3.5 is real?
Ignore secondary coverage and screenshots. A model is real when the vendor says so on its own surfaces. Check these, in order:
- blog.google/products/gemini — Google's official Gemini hub. A real launch gets a post here with the official name.
- The Gemini app and the Gemini API / Vertex AI model list — a callable model ID and a model card are the definitive signal. If you can select it in the app or call it by ID, it shipped.
- An official model or system card with capabilities and intended-use documentation — this is where any real benchmark numbers will first appear.
Until those exist, “Gemini 3.5” is a name in leaks. Treat any spec you see as unverified, and check the primary sources above before you plan around it.
FAQ
Did Google announce Gemini 3.5 at I/O 2026?
Not confirmed as of this writing (the morning of May 19, 2026). Google I/O 2026's keynote is scheduled for today but has not delivered an official “Gemini 3.5” name or model. Watch blog.google/products/gemini for the authoritative announcement — this page updates in place when one lands.
Is Gemini 3.5 Pro available in the API?
No. There is no gemini-3.5-pro model ID in the Gemini API or Vertex AI as of May 19, 2026. The current flagship you can call is Gemini 3 (Pro tier).
When will Gemini 3.5 be released?
Unknown. Google has not published a release date or confirmed the product. Anyone citing a specific date is repeating speculation, not an official announcement.
What is “Cappuccino” for Gemini?
It is an internal codename for a candidate Gemini checkpoint, reported by secondary outlets around May 15–17, 2026. It is not confirmed to be “Gemini 3.5,” not confirmed to ship under any public name, and not confirmed to ship on any date.
Are the Gemini 3.5 benchmarks I saw real?
No published Gemini 3.5 benchmark is verifiable, because there is no official model card or system card. Treat any circulating “Gemini 3.5” benchmark table as unverified until Google publishes one.
Gemini 3 vs Gemini 3.5 — which should I use?
Use Gemini 3. “Gemini 3.5” is not a model you can use — it is not released. Gemini 3 (Pro) is the current shipping flagship for reasoning and coding.
Where is the official Gemini 3.5 announcement?
There isn't one. The authoritative place it would appear is blog.google/products/gemini, alongside a callable API model ID and a model card. None of those exist for “Gemini 3.5” as of this writing.
Hiring developers who can move with the model field?
If your team is shipping on a fast-moving model landscape — keeping the model layer swappable, evaluating new releases against what's actually in production, and not getting whipsawed by every pre-keynote rumor — you want engineers who have done exactly that. codersera.com/hire matches you with vetted remote developers experienced with LLM integration, evaluation, and self-hosting, so you can adopt the right model on evidence rather than hype.