Saad Ullah
⬤ Google has quietly added Gemini 3 Deep Think to the Gemini app - Ultra subscribers now reach the firm's most powerful reasoning engine. The move prompts the same question everywhere - did Google give the model more time to think or did it rebuild the stack? The official line is terse - Gemini 3 Deep Think solves math, science and logic questions that routinely defeat the best public models.
⬤ The headline figures show large margins. With no external tools the system scores 41 % on Humanity's Last Exam. When it may write and run code, the figure on ARC-AGI-2 climbs to 45.1 %. Google traces the jump to a parallel reasoning scheme refined from the Gemini 2.5 Deep Think family. Earlier checkpoints already took gold at the International Mathematical Olympiad and at the ICPC World Finals, evidence that the lineage already handles exacting, real world tasks.
⬤ That simultaneous search is the claimed differentiator. Ultra users activate it - choosing “Deep Think” in the prompt bar. Outside observers still argue about internals - yet the drop aligns with Google's wider plan to pitch Gemini as the safe choice for high stakes analytic work.
⬤ For anyone tracking GOOG, the numbers carry weight. Outsize benchmark results strengthen the narrative that Google keeps pace in the capability race, a race that now moves share prices. Each new reasoning tier raises the odds that large firms will standardise on Gemini and it secures Google's berth in the contest for primacy in artificial intelligence.
Saad Ullah
