Google Gemini 3 'Deep Think' mode and benchmark claims

Google's Gemini 3 launches Deep Think mode.
Key insights
Benchmark lead on ARC‑AGI‑2
Gemini 3's Deep Think mode is reported to achieve 45.1% accuracy on the ARC‑AGI‑2 benchmark, which posts compare to GPT‑5.1's 17.6%, cited as a 2.5× gap. The release also claims near‑top scores on GPQA Diamond and a 41.0 score on the Humanity’s Last Exam (HLE) benchmark. These numeric claims are presented as measured improvements versus prior model baselines.
Parallel 'multi‑hypothesis' reasoning feature
Deep Think mode is described as supporting parallel or multi‑hypothesis reasoning rather than linear step‑by‑step chains. The mode reportedly explores multiple solution paths in parallel, enabling more detailed and deliberative answers for hard reasoning tasks. Posts describe this as a core architectural/behavioral difference in Gemini 3's Deep Think reasoning workflow.
Availability limited to Ultra tier users
Deep Think mode is reported as available to 'Ultra' users. One post describes UI access: select 'Deep Think' on the left and choose 'think' on the right. The rollout detail indicates an initial gated availability tied to a premium user tier.
Claims framed as competitive move vs OpenAI
Posts frame the Deep Think launch as a competitive step against OpenAI, emphasizing benchmark dominance in high‑difficulty tests and calling out direct comparison to GPT‑5.1 numbers. The narrative positions Deep Think as a major Google advancement in deep reasoning capability.