Gemini 3 Deep Think Gets a Major Upgrade: 84.6% on ARC-AGI-2 and 18 Unsolved Problems Cracked

Google Pushes the Reasoning Frontier with Deep Think

On February 12, 2026, Google announced a major upgrade to Gemini 3 Deep Think, its specialized reasoning mode designed to tackle the hardest problems in science, research, and engineering. The update delivers record-breaking benchmark scores and, more importantly, demonstrates practical problem-solving capabilities that extend well beyond standardized tests.

Deep Think is not a separate model but rather a reasoning mode within the Gemini 3 family. When activated, it allocates significantly more computation time to work through complex, multi-step problems. This upgrade refines that process with improvements developed in close partnership with scientists and researchers who face real-world challenges where data is messy, solutions are ambiguous, and problems often lack clear guardrails.

Benchmark Performance That Demands Attention

The numbers from this upgrade are striking across multiple domains. On ARC-AGI-2, the abstract reasoning benchmark designed to test general intelligence capabilities, Deep Think achieved an unprecedented 84.6 percent, verified independently by the ARC Prize Foundation. For context, the standard Gemini 3 Pro scores 31.1 percent on the same test, and Claude Opus 4.6 reaches 68.8 percent. On the original ARC-AGI-1 benchmark, Deep Think essentially saturated the test at 96 percent.

On Humanity's Last Exam, a benchmark specifically designed to push the limits of frontier models, Deep Think scored 48.4 percent without tools. In competitive programming, it achieved a Codeforces Elo of 3,455, placing it among the world's top competitive programmers. On GPQA Diamond, a graduate-level science benchmark, it scored 93.8 percent, narrowly edging out GPT-5.2 Pro at 93.2 percent.

Deep Think also earned gold-medal-level results on the written sections of both the 2025 International Physics Olympiad and the 2025 International Chemistry Olympiad.

Real-World Problem Solving Beyond Benchmarks

What distinguishes this upgrade from a typical benchmark improvement is the evidence of practical scientific impact. Google reports that Deep Think solved 18 previously unsolved research problems and disproved a decade-old mathematical conjecture from 2015.

In one notable case, a mathematician at Rutgers University used Deep Think to review a highly technical mathematics paper. The model successfully identified a subtle logical flaw that had previously passed through human peer review unnoticed. This represents a qualitatively different kind of capability from answering exam questions: the ability to catch errors that trained experts miss.

Another practical demonstration involves converting hand-drawn sketches into 3D-printable objects. Deep Think can analyze a drawing, model the complex shape mathematically, and generate a file ready for 3D printing. While this may sound like a novelty, it illustrates the model's ability to bridge abstract understanding and concrete engineering output.

How Deep Think Compares to the Competition

The AI reasoning landscape as of February 2026 is increasingly specialized. Each frontier model excels in different domains:

Deep Think dominates in abstract reasoning and scientific problem-solving. Claude Opus 4.6 remains the leader for real-world coding tasks and agentic workflows, scoring 80.8 percent on SWE-bench Verified. GPT-5.2 holds strong in mathematical reasoning. The takeaway is that the frontier is no longer defined by a single model but by a collection of specialized capabilities.

Availability and Access

The upgraded Deep Think is available through two channels. Individual users can access it through the Gemini app with a Google AI Ultra subscription. For the first time, Deep Think is also available via the Gemini API, targeting researchers, engineers, and enterprise users. Google is providing early access through an application process for API users.

Google has not disclosed specific pricing for Deep Think API calls, though Gemini API pricing generally remains competitive. The AI Ultra subscription that includes app access is priced at $24.99 per month.

What This Means for AI Reasoning

The Deep Think upgrade represents a meaningful step in a trend that has been building throughout 2025 and into 2026: the emergence of reasoning modes that trade speed for depth. Rather than trying to answer every query as quickly as possible, these systems allocate extended computation time to work through genuinely difficult problems.

The practical implications are significant for researchers and engineers. A model that can review papers for logical flaws, solve previously intractable research problems, and bridge the gap between abstract reasoning and physical engineering output is not just a benchmark achievement. It is a tool that could meaningfully accelerate scientific work.

However, it is worth noting that Deep Think's strengths are concentrated in structured reasoning tasks. For everyday conversational AI use, code generation, or creative writing, the standard Gemini 3 Pro or competing models may be more appropriate and faster.

Conclusion

Google's Gemini 3 Deep Think upgrade sets new standards in AI reasoning with its 84.6 percent ARC-AGI-2 score and practical scientific applications. It is best suited for researchers, engineers, and scientists who need an AI that can engage with genuinely difficult problems at a deep level. For users whose work involves complex analysis, mathematical reasoning, or scientific research, Deep Think is now the strongest option available. For general-purpose AI tasks, the standard frontier models remain the better choice.