Gemini 3.5 Pro: Release Tracker, Leaks, and What's Actually Confirmed

Everyone has been saying Gemini 3.5 Pro ships "today" since last week. And the week before that. And in July.

The problem is that "today" has become a permanent state. LM Arena screenshots, a thread with 4,000 retweets, aggregator blogs rewriting the same rumor under a new headline. None of that is Google talking. Meanwhile, people are putting off architecture decisions waiting for a model nobody can put a date on.

This post is a tracker, not a prophecy. We're going to separate three things: what Google has confirmed in writing, what's a leak with no reproducible data, and how to build a 20-line watcher that pings you the minute gemini-3.5-pro lands in the API — without depending on anyone on X.

TL;DR

• What this is: a Gemini 3.5 Pro launch tracker, with an explicit line between official fact and rumor.
• Status as of August 8, 2026: not released. It doesn't show up in the Gemini API changelog or in the models endpoint.
• Confirmed by Google: the model exists, it was announced at I/O in May, and it's in internal use. Nothing beyond that.
• Rumor: 2M-token window, native Deep Think, dates from August 12 to 25, internal codename.
• What to do: read the watcher section and stop refreshing your timeline.

The context: why Gemini 3.5 Pro turned into a soap opera

The timeline helps you see how big the delay is.

On February 19, 2026, Google released Gemini 3.1 Pro, which is still the top of the Pro family: 1M input context, 64k output, 77.1% on ARC-AGI-2, 94.3% on GPQA Diamond and 80.6% on SWE-Bench Verified.

On May 19, 2026, at I/O, Google announced the 3.5 family in the post Gemini 3.5: frontier intelligence with action. Flash shipped the same day, with 76.2% on Terminal-Bench 2.1 and 83.6% on MCP Atlas. About Pro, the text was specific: it was already in internal use and would arrive "next month." In other words, June.

June went by. July went by. On July 21, Google released three models at once — gemini-3.6-flash, gemini-3.5-flash-lite and a Cyber variant — and Pro wasn't on the list, as TechCrunch put right in the headline. If you want the numbers from that release, they're in the post on Gemini 3.6 Flash.

The most cited reason from a serious source came from Bloomberg on July 16: coding performance below Google's own internal targets. Follow-up reporting talks about scrapping the previous training run and redoing pre-training from scratch — the kind of thing that takes months, not weeks.

That's the part that matters if you build things: while the market argues about a launch date, the dev's real problem is having a system that doesn't break when the model changes. It's exactly that kind of boring decision — provider abstraction, regression evals, rollback plan — that we put on the table every week, live, at Clã Beer and Code. It's a paid subscription, it's devs building with devs, and there are no recorded lessons.

What Google has officially confirmed about Gemini 3.5 Pro

Short list. Every line here has a Google URL behind it.

1. Gemini 3.5 Pro exists and was announced. It's in the official I/O post from May 19, 2026. No leaker made it up.
2. It was in internal use in May. Google's text says "already being used internally".
3. The original target was June 2026. Also from the official post. Already blown.
4. It's still listed as "coming soon." The DeepMind models page still shows Gemini 3.1 Pro as the current Pro, with 3.5 Pro marked as next.
5. It's not in the public API. The Gemini API changelog has no gemini-3.5-pro entry. The latest entries are from July 21 (gemini-3.6-flash, gemini-3.5-flash-lite) and July 30 (Gemini Robotics ER 2).

That's it. No official benchmark. No pricing. No context window. No date.

What's a leak and nothing more: dates, benchmarks and codename

Everything below circulates as if it were fact. Not one item has confirmation from Google.

• 2M-token window. Repeated by practically every aggregator. The origin is always "sources," never a model card.
• Native Deep Think. Plausible, since Deep Think has existed since Gemini 3, but not confirmed for 3.5 Pro.
• Internal codename "Cappuccino." Showed up in leaks from mid-July.
• Sightings. The model supposedly appeared on LM Arena and vanished within 44 minutes, labeled as gemini-3.5-flash. It was also supposedly seen running inside Antigravity in a stealth test.
• Dates. There's a new one every week: July 17, July 24, the first week of August, August 12 to 14. On Polymarket, as of July 30 the bets were concentrated on August 25 (41%) and August 15 (39%) — a prediction market is a signal, not a source.
• Leaked benchmarks that contradict each other. One leak says the stealth-test version beat Claude Fable 5 on a front-end task. Another says it trails Fable 5 and GPT-5.6 on reasoning and code. Neither publishes a dataset, a prompt or a methodology.

Rule of thumb: if the claim doesn't come with a reproducible number and a link to the model card, it's noise. An arena screenshot isn't a benchmark. It's a screenshot.

Tracker: how to detect the launch without depending on a thread on X

This is the useful part. The Gemini API's ListModels endpoint is the fastest source of truth there is: a model shows up there the moment it becomes available, often before the post on Google's blog starts making the rounds.

Step 1: see what's available right now

curl -s "https://generativelanguage.googleapis.com/v1beta/models?key=${GEMINI_API_KEY}&pageSize=200" \
  | jq -r '.models[].name' \
  | grep -i 'gemini-3'

If models/gemini-3.5-pro isn't in that list, the model doesn't exist for you. Period. It doesn't matter how many screenshots you've seen.

Step 2: a watcher that runs on its own

Instead of checking by hand, keep a snapshot and diff against it. That way you get alerted about any new model, not just the one you're waiting for — and sometimes the interesting one is the one nobody announced.

#!/usr/bin/env bash
set -euo pipefail

SNAP="${HOME}/.cache/gemini-models.txt"
mkdir -p "$(dirname "$SNAP")"

atual=$(curl -sf "https://generativelanguage.googleapis.com/v1beta/models?key=${GEMINI_API_KEY}&pageSize=200" \
  | jq -r '.models[] | select(.supportedGenerationMethods[]? == "generateContent") | .name' \
  | sort -u)

if [ ! -f "$SNAP" ]; then
  printf '%s\n' "$atual" > "$SNAP"
  echo "snapshot inicial gravado com $(wc -l <<<"$atual") modelos"
  exit 0
fi

novos=$(comm -13 "$SNAP" <(printf '%s\n' "$atual") || true)

if [ -n "$novos" ]; then
  echo "modelos novos na Gemini API:"
  printf '%s\n' "$novos"
  printf '%s\n' "$atual" > "$SNAP"
fi

Drop it in cron every 15 minutes and send the output to the channel you actually read:

*/15 * * * * /usr/local/bin/gemini-watch.sh | mail -s "Gemini API" voce@exemplo.com

Swap mail for a curl to a Slack, Discord or Evolution API webhook. The point is the same: the notification comes from the endpoint, not from an influencer.

Step 3: confirm before you celebrate

A model showing up in the list still doesn't mean it's available to you. Before touching any production code, check three things:

• Does it actually run? Make a minimal generateContent call. Plenty of models show up in preview behind an allowlist and return a 403.
• What does it cost? Preview pricing is often different from GA pricing, and the end-of-month bill is unforgiving.
• Is there a model card? Without a published model card, you have no verifiable benchmark and no stated usage limits. It's still a rumor, just one with an API name.

What to do while Gemini 3.5 Pro isn't out

The right answer isn't to wait. It's to get your system ready to swap the model with one line of config.

If you're on Laravel, this is practically free with Prism:

// config/ai.php
return [
    'principal' => env('AI_MODEL', 'gemini-3.1-pro'),
    'fallback'  => env('AI_MODEL_FALLBACK', 'gemini-3.6-flash'),
];

namespace App\Services;

use Illuminate\Support\Facades\Log;
use Prism\Prism\Enums\Provider;
use Prism\Prism\Prism;

class GeradorDeResposta
{
    public function gerar(string $prompt): string
    {
        $modelos = [config('ai.principal'), config('ai.fallback')];

        foreach ($modelos as $modelo) {
            try {
                return Prism::text()
                    ->using(Provider::Gemini, $modelo)
                    ->withPrompt($prompt)
                    ->asText()
                    ->text;
            } catch (\Throwable $e) {
                Log::warning('modelo indisponível, caindo pro próximo', [
                    'modelo' => $modelo,
                    'erro' => $e->getMessage(),
                ]);
            }
        }

        throw new \RuntimeException('nenhum modelo Gemini respondeu');
    }
}

With that in place, launch day becomes an AI_MODEL=gemini-3.5-pro in your .env and an eval run. It doesn't become a refactor.

And the eval run is the part almost everyone skips. Have a set of 30 to 50 real cases from your domain, with expected answers, running before deploy. A new model isn't automatically better at your problem — only at the vendor's benchmarks. I've seen a model swap improve reasoning and break JSON formatting in the same commit.

Limitations and things to watch out for

• This post has a date on it. It was written on August 8, 2026. If you're reading this later, run the curl from step 1 before believing anything here.
• ListModels isn't infallible. A model can be released on Vertex AI for enterprise customers first and only later on the public Gemini API. The watcher covers the public API; if you're a Vertex customer, also monitor the Vertex release notes.
• Preview breaks. A preview model changes behavior without warning and can be deprecated. Don't put a preview alias on a critical path without a fallback.
• Careful with the key in cron. Don't leave GEMINI_API_KEY hardcoded in the script or in the crontab. Read it from a file with 600 permissions or from a secrets manager.
• Don't migrate out of FOMO. Switching models costs evals, costs QA and costs risk. If 3.1 Pro handles your use case today, it'll keep handling it after the launch.

Quick FAQ

Is Gemini 3.5 Pro out yet? As of August 8, 2026, no. It's not in the Gemini API changelog and it doesn't respond on the public models endpoint. The current Pro available is still Gemini 3.1 Pro.

Is the 2M-token window confirmed? No. It's the most repeated rumor and the least supported. Google hasn't published any context spec for 3.5 Pro. 3.1 Pro has 1M input and 64k output, and those are actually documented.

What's the Gemini 3.5 Pro release date? There is no official date. Google only committed to "June 2026" in the I/O post, and that date has already slipped. Every other date going around comes from a leak or a prediction market. Treat any specific date as a bet, not a schedule.

Why has it been delayed so long? The most reliable reporting, from Bloomberg on July 16, 2026, points to coding performance below internal targets. There are reports of retraining from scratch. Google hasn't commented publicly on the reason.

What do I use in the meantime? Gemini 3.1 Pro for heavy reasoning tasks and Gemini 3.6 Flash when latency and cost matter more. Both are GA and documented. And get the fallback ready in your code.

Conclusion

Gemini 3.5 Pro will ship. Probably in August, possibly in September, and anyone who gives you a date with certainty is guessing with confidence.

What doesn't change is the method: official fact has a vendor URL and a model card. Rumor has a screenshot. While the market burns energy guessing dates, the dev who takes AI seriously spends the same 20 minutes setting up a watcher on the endpoint and a model fallback in the code — and on launch day flips an environment variable while everyone else is still reading threads.

That's not hype. It's engineering.