We’ve upgraded our specialized reasoning mode Gemini 3 Deep Think to help solve modern science, research, and engineering challenges – pushing the frontier of intelligence. 🧠 Watch how the Wang Lab at Duke University is using it to design new semiconductor materials. 🧵

The latest Deep Think moves beyond abstract theory to drive practical applications. It’s state-of-the-art on ARC-AGI-2, a benchmark for frontier AI reasoning. On Humanity’s Last Exam, it sets a new standard, tackling the hardest problems across mathematics, science, and engineering — making it a genuine collaborator for heavy-duty analysis. It achieved an Elo of 3455 on Codeforces, demonstrating the ability to solve complex, real-world coding tasks - while earning gold medal-level results on the written portion of the 2025 Physics and Chemistry Olympiads.

The upgraded Deep Think mode is rolling out now in the @GeminiApp for Google AI Ultra subscribers. For scientific researchers and developers, we’re opening a Vertex AI Early Access Program for the API. Start discovering →

84.6% on ARC-AGI 2, and it's only February of 2026. Google cooked.

Model companies waiting for their competitor to release a new model so they can release theirs a day later and steal the news cycle.

damn that is not a little upgrade guys

brutal frame mog for gptcels holy cortisol spike for opuscels giga lifefuel for geminicels over for arc-agi 2 benchmarkcels never began for "the wall" copers

Damn, they mogged so hard!!

crazy benchmarks!!!!!

other model just pushing their agentic capabilities ... google push science you have my respect ...

Looks like @grok has some catching up to do! I wonder what Grok 4.2 will fare in this match-up?

everyone was watching the anthropic vs openai show and google just quietly posted the highest score on the board

@Vicrom1509 Why are you not giving deep think model to pro users? 😞

no for pro users? what's wrong with the AI world and its subscription plan right now?

Gotta love these rate limits despite me paying already 250$ a month

84.6% on ARC-AGI 2, and it's Only February of 2026... Google cooked...

Google trains and runs Gemini (incl. Gemini 3 Pro) primarily on its own TPUs, not Nvidia GPUs. Why this matters: • TPUs = custom AI chips (ASICs) built in-house since 2013 • Higher efficiency, lower cost, and better scaling for Google’s workloads • Independence from Nvidia pricing + supply constraints • DeepMind co-designs TPUs → tight model–hardware optimization

The new best is the new worst it will ever be. Looks like a great model!

The real frontier isn’t AI replacing scientists — it’s scientists expanding what they can attempt.

its all making sense now

Google is always going to win the AI race. Not even close.

My cousin used Gemini to create a serum that would turn his eyes gray. He went blind.

That’s not “AI writing essays.” That’s AI accelerating science.

How do I use deep think in the Gemini app?

Other model just pushing their agentic capabilities... Google push Science. You have My Respect...

Amazing progress in a few months, imagine what models we don't get to see, that they use internally!

I’d love to see a frontiermath benchmark!

"Please give Google your cutting edge tech designs" said no one sane, ever.

is the limit still 10 a day?

AI that moves from theory to practical application is where the real disruption happens. Semiconductor design is just the beginning. I can't believe how much progress has been happening with these models.

The semiconductor materials design use case is a perfect demo — that's exactly where deep reasoning shines over pattern matching. 84.6% on ARC-AGI-2 is wild though, wasn't this benchmark supposed to be hard for years?

Google and every single AI Company are going to cause Planetary Extinction

This is amazing. As I said earlier, Ai wil start evolving

Using AI to design semiconductors is the recursive loop everyone predicted but nobody's shipped at production scale. The real question is whether Deep Think compresses the 18-month fab iteration cycle. If Duke's lab goes from simulation to tape-out in weeks instead of months, that's the benchmark worth tracking.

I don't see the reason to continue paying for the Gemini_Pro version. It is significantly inferior to its competitors at a similar price. You could offer 2-3 requests per day to Deep Think.

does this concern internal Gemini Deep Think model not available to public?