Second, some coverage points to alleged Google AI Studio A/B tests or hidden checkpoints. Geeky Gadgets describes leaked AI Studio testing under the working title “Gemini Advance,” with claimed gains in instruction-following and creative tasks but also inconsistencies in visual outputs and other areas [8]. YouTube summaries are even less consistent: one says the alleged Gemini 3.5 test performed worse than Gemini 3.0 Pro, while another claims a 40% boost in coding and reasoning [
11][
13].
That inconsistency is the key issue. One YouTube snippet explicitly warns that anonymous benchmark screenshots and forum claims were not verified or reproducible [16]. In other words, the leaks may be real signals of experimentation, but the available sources do not prove a released model, a stable benchmark, or a launch date.
There is no evidence-backed date. Some rumor posts floated early-2026 timing, including February or March expectations [6][
10]. But the official Google listing cited here still shows Gemini 3.1 preview models and no Gemini 3.5 entry [
1].
The only quantified timing signal in the available sources is a Manifold prediction market. It showed 34% odds for release before June 2026, 44% before July, 60% before August, and 67% before September [2]. That points to a speculative mid-2026 expectation among bettors, not a Google roadmap.
The market’s own rules are a useful reality check: it counts a release only if the model is accessible without individual vetting or invitation, and it excludes closed betas, invite-only access, researcher programs, A/B tests, and anonymous benchmark appearances [2]. By that standard, the current leak claims would not count as a public Gemini 3.5 launch.
A credible public release should appear in an official Google channel, such as Gemini API documentation or a Google product announcement. In the available official documentation, Gemini 3.1 preview models are listed and Gemini 3.5 is not [1].
Until that changes, treat these as weak evidence rather than confirmation:
Those signals can be useful to monitor, but they are not the same as a release note, model card, API listing, or public product rollout.
Gemini 3.5 is not confirmed as a publicly released model, and there is no official release date in the available Google documentation [1]. A summer 2026 window is a reasonable way to describe current speculation because prediction-market expectations cluster around June through August, but it should be labeled speculation unless Google publishes an official Gemini 3.5 listing or announcement [
2].
- Google’s Gemini 3.5, revealed through leaks, shows advancements in instruction-following and creative tasks but faces inconsistencies in visual outputs and underwhelming performance in specific areas. ... Google Gemini 3.5: Leaked Insights Google’s Gemini...
January 13, 2026 ... Leaked benchmark data from an internal checkpoint codenamed ‘Snow Bunny’ reveals that Google Gemini 3.5 can generate 3,000 lines of executable code from a single prompt and scores 80% on difficult logic tests — a 25-percentage-point gap...
According to leaked data from late December and early January, Google is already deep into testing Gemini 3.5 . The internal code name? Snow Bunny . ... Leaks suggest Google is planning to roll out Gemini 3.5 publicly by February 2026 — possibly sooner if t...
Jan 20, 2026 (0:08:24) Gemini 3.5 just leaked through Google's AI Studio A/B testing and early results show it's actually performing WORSE than the current Gemini 3.0 Pro. Meanwhile, OpenAI's Sam Altman confirms GPT-5.3 is in development, and DeepSeek drops...
Google is secretly testing Gemini 3.5, and early leaks show a massive 40% performance boost in coding and reasoning. ... {ts:0} New Google Gemini 3.5 leaks are insane. Google is secretly testing Gemini 3.5 {ts:6} right now and people are getting access thro...
… {ts:250} official release. There are anonymous benchmark screenshots floating around. AI focused forums are going crazy. But {ts:256} here's the problem. None of this is verified. We can't confirm the sources. We can't replicate the tests. It's all {ts:26...
Comments
0 comments