gemini 3.5 pro leak: what we know and how x users are testing it

release date, context window, a new reasoning mode, and the real reason for the delay – everything the gemini 3.5 pro leaks point to, plus what x users are already building with it. here's what's confirmed and what's rumored

what's rumored

• release around july 17. multiple accounts point to mid-july, though no date is official

• 2m token context window. this would be roughly 2x anthropic's current 1m ceiling (fable 5, sonnet 5, opus 4.8), and among the largest of any major frontier model

• a new "deep think" reasoning mode aimed at logic and math

• stronger autonomous agent support for multi-step tasks
• a fresh base model. one leaker claims google scrapped the old base and started over
• mixed verdicts on where it lands. some leaks claim stronger zero-shot performance than claude fable 5 in private eval rings. others claim the opposite – that even with a new base it still won't beat fable 5 or gpt 5.6 on the hardest agentic and long-horizon work

what we know

the facts, mostly traced back to a business insider report and google's own i/o comments:

• it slipped from june to july. google previewed the model at i/o in may 2026, where sundar pichai said it would arrive "next month." the june window closed with no public launch
• the delay is about feedback, not a rewrite of the story. per business insider, google pushed the date back to gather more real-world use cases from early testers and keep tuning the model
• it's already in limited hands. early access exists through a vertex ai enterprise preview, google's antigravity platform, and the benchmarking site lmarena. broad public access hasn't opened
• it's aimed at long-horizon and agentic work. the model is positioned around long-context reasoning and running autonomous agents, not just faster chat
• flash 3.5 feedback got folded in. google reportedly took criticism that the earlier flash model burned through tokens too quickly and worked on token efficiency for the pro tier
• google is staying quiet. a spokesperson declined to comment on the schedule, and there are no official evals or benchmark numbers yet

what people are generating

the bulk of the public testing comes from two accounts, harshith (@harshithlucky3) and @ggg78g89, both running single-shot generation prompts and posting the outputs. worth noting: the model tags carry asterisks even from the testers themselves, since the arena checkpoints are stealth and nobody can confirm which model actually produced what. with that caveat, here's what they're buildingthree.js / 3d scenes – the model's most-shared category is fully working 3d scenes in a single html file:

• a futuristic maglev bullet train speeding through a transparent glass vacuum tube (posted as a head-to-head against fable 5 xhigh)

• a voxelized human heart that animates the voxels to expand and contract like a real heartbeat (run against a "3.6 or 4 flash" checkpoint)

• a floating steampunk island

• a cinematic moon scene – astronaut, earthrise on the horizon, dusty footprints, a planted flag, harsh sunlight, black-space mood – built as one working html file

• an animated american flag where stars appear one by one to mark states joining the union over time, posted as a july 4th / 250th tribute

inline svg, technical precision – the community's benchmark is single-prompt svg, and the "precision" prompts are meant to stress geometry and detail:

• a minimal isometric card swiping machine (the original viral arena prompt)

• a bmw m4 cs side view (from an early "3.6 / 4 flash" checkpoint test)

• an analog wristwatch reading 10:08:42, with a date window on 5, full 60-minute tick marks, metallic gradients and no text labels – the poster claimed this one edged out fable 5

inline svg, narrative vignettes – a second flavor of svg prompt tests emotional and compositional range rather than mechanical accuracy:

• a stray dog and a street kid sharing a bun under a tiny shop awning in the rain, their shadows merging into one (posted next to a gemini 3.1 pro output for comparison)

• a mother quietly fixing her child's school uniform before sunrise, polished shoes waiting by the door

• a tense scene of a mouse stealing cheese from a mousetrap while a cat sleeps nearby, lit by moonlight (posted as the model "mogging" fable 5 and glm-5.2)

functional apps and games – a few testers pushed past static scenes into interactive builds:

• a web-based, playable 3d subway-surfers-style game generated as roughly 800 lines of html – a city with streets, crossings, houses, moving cars, pedestrians, working traffic lights, skyscrapers and shops, with the model noting up front it would deliver a low-to-mid-poly prototype

almost everything being generated is visual and single-shot – three.js worlds and svg scenes that either show off geometric precision or emotional composition. that's consistent with the "frontend mogging" reputation forming around the model, and it's also why the picture stays incomplete. these prompts test spatial and graphical reasoning, not the long-horizon agentic and debugging work that the other set of leaks says the model is weaker at

conclusion

even through the noise, the early outputs show a model that's clearly capable and competitive with the current frontier. and if the 2m token context window holds up, that alone could push it into the lead – twice the room of anything anthropic ships today, on a model already looking strong on frontend and visual reasoning. for now it's still leaks and stealth checkpoints. we'll wait for the release and test it ourselves