Gemini 3.5 Pro: is it out yet? What we know (2026)

Is Gemini 3.5 Pro out yet?

No. I check this a lot, because "Gemini 3.5 Pro" is one of the most-searched AI queries right now, and the search results are full of confident-sounding spec sheets for a model that doesn't have a public page.

Here's what three separate Google-owned properties actually say as of July 2026:

• The Gemini Pro model page shows the current flagship as Gemini 3.1 Pro with a "3.5 Pro coming soon" badge sitting right above it.
• The Gemini API model list has no gemini-3.5-pro entry anywhere, in Stable or Preview. The newest Pro you can call is gemini-3.1-pro-preview.
• The API pricing page has a Gemini 3.1 Pro row and a Gemini 3.5 Flash row, but no 3.5 Pro price.

So the short version: the 3.5 generation is real and partly shipped, but the Pro member of it isn't. That distinction trips people up, so here's the whole family at a glance.

What Google has actually confirmed

I'll stick to what's on Google's own pages, because that's the only ground worth standing on here.

The 3.5 series was announced on May 19, 2026 at Google I/O, in a post bylined by DeepMind CTO Koray Kavukcuoglu alongside Jeff Dean, Oriol Vinyals, and Noam Shazeer. The pitch for the whole line is "frontier intelligence with action," meaning it's tuned for agentic workflows and coding, not just chat. On the same page, Google wrote that 3.5 Pro is "already being used internally" and would "roll it out next month." That post is dated May, so "next month" meant June. It's now well past that.

Only Gemini 3.5 Flash shipped at launch, and Google says it already beats the older Gemini 3.1 Pro on several agentic and coding tests. Those Flash numbers are the only 3.5-generation benchmarks Google has published, so they're the best floor for guessing where 3.5 Pro will land:

The standout Flash results Google highlights are Terminal-bench 2.1 at 76.2% for agentic terminal work, MCP Atlas at 83.6% for tool calling, and CharXiv Reasoning at 84.2% for reading complex charts, plus a claim of being "4 times faster than other frontier models" on output speed. One more clarification worth making: the shipped "Deep Think" reasoning mode is Gemini 3.1 Deep Think, not a 3.5 Pro variant. Google hasn't announced a 3.5 Pro Deep Think, so don't treat one as confirmed.

Why is Gemini 3.5 Pro taking so long?

Google hasn't given a blow-by-blow, but the shape of the delay is clear from its own "coming soon" status plus what the developer community is saying out loud. The frustration is real, and it's the loudest thread about 3.5 Pro right now.

The mood is disbelief that Google went from front-runner to months behind:

"Near the end of 2025, Google was on top - best LLM (Gemini 3 Pro), best video model (Veo 3.1), and best image model (Nano Banana). They seemed unstoppable. And then, they stopped. Why?"

pants2, Hacker News

And developers keep tying the delay back to one specific weakness, reliability:

"Alphabet Inc.'s Google is months behind schedule on delivering Gemini 3.5 Pro, its most powerful flagship AI model, because the company has been taking time to try to improve its capabilities, particularly in coding... Seems like deepmind needs to get more internal usage"

htrp, Hacker News

That "improve reliability" theme matters if you're evaluating any Gemini model for real work, so it's worth pulling apart before you commit.

What you can actually use today: Gemini 3.1 Pro

While you wait, Gemini 3.1 Pro is the flagship, and it's not a consolation prize. On Google's own performance table it leads on GPQA Diamond (94.3%), Humanity's Last Exam (44.4%), ARC-AGI-2 (77.1%), and LiveCodeBench Pro (2887 Elo), and it trades the top spot on SWE-Bench Verified with Claude Opus by a hair (80.6% to 80.8%). The core specs:

Specs from the Gemini 3.1 Pro model page. It's available in the Gemini app, Google AI Studio, the Gemini API, and Google Antigravity.

Gemini 3.1 Pro and 3.5 Flash pricing

This is where the "which one" decision usually gets made, so here's the real API pricing (per million tokens, standard tier). Pro is the only family with pricing that jumps at the 200k-token mark.

Numbers from the Gemini API pricing page. Cheaper Batch and Flex tiers roughly halve those rates, and a Priority tier costs more. One gotcha to budget for: grounding with Google Search is free for the first 5,000 prompts a month across the whole Gemini 3 family, then $14 per 1,000 queries after that.

If you'd rather pay a flat consumer price than meter tokens, Google renamed its plans after I/O:

Live prices from Google AI plans. For a fuller cost breakdown, including how these stack up against Gemini 3 pricing and the best Gemini alternatives, those two guides go deeper than I can here.

What real users say about Gemini Pro

Benchmarks are one thing, daily use is another, and the developer verdict on the Gemini Pro line is genuinely split. The wins are real: several people treat 3.x Pro as their daily driver for its speed and lower everyday wrongness than ChatGPT.

"In my experience Gemini 3.0 pro is noticeably better than chatgpt 5.2 for non-coding tasks. The latter gives me blatantly wrong information all the time, the former very rarely."

tempestn, Hacker News

But the recurring complaint, and the exact thing the 3.5 Pro rebuild is meant to fix, is confident hallucination:

"Use ChatGPT, and it's ask a question, get an answer, done... But use Gemini? You get burned by the hallucinations Gemini casually spews out on a daily basis. You develop a habit of fact checking every single answer."

jdw64, Hacker News

And for agentic coding specifically, the sentiment drops further, with reports of doom-loops rather than raw-intelligence gaps:

"My sense is that the Gemini models are very capable but the Gemini CLI experience is subpar compared to Claude Code and Codex... it can get confused, fall into doom loops, and generally lose the plot."

foruhar, Hacker News

That's the buyer tension in one line: the benchmarks say try it, the real world says babysit it. Which is exactly why, for anything customer-facing, I care less about the model number and more about the layer on top.

The model is the engine, not the car

Here's the reframe I'd leave you with, especially if you found this post while deciding which model to build customer support on. Frontier models leapfrog each other every few weeks. Gemini 3.1 Pro leads today, 3.5 Pro will leapfrog it, GPT and Claude will answer, and round it goes. If your support automation is welded to one model, every one of those releases is a migration project.

The part that actually decides whether an AI support agent is any good isn't the base model, it's everything wrapped around it: whether it's trained on your knowledge, whether you can set guardrails on what it's allowed to say, and whether you can test it before it touches a real customer. I've spent enough time watching a confident-sounding bot give a wrong answer to know that the hallucination worry above is the whole ballgame for support. The fix isn't a better model number, it's a system that simulates against your real past tickets first and only auto-replies where it's proven safe.

Try eesel for AI support that isn't locked to one model

That system is what eesel is. It's an AI agent for customer support that runs on frontier models like Gemini under the hood, plugs into your existing helpdesk (Zendesk, Freshdesk, Gorgias, and more), and trains on your past tickets and help center. So when the next Gemini, GPT, or Claude ships, you get the upgrade without re-architecting anything.

The differentiator that maps straight to the hallucination fear above: eesel simulates on thousands of your historical tickets before it goes live, so you see the resolution rate and exact answers on real questions first, then dial up automation only where it's safe. It's free to try and sets up in minutes, no model babysitting required.