Google DeepMind’s Deep Think Dominates Eight Language Olympiads and Solves Four AI Challenges

Deep Think, the AI system from Google DeepMind, has achieved record‑high scores across a series of eight language‑specific competitions. In July 2025 the Gemini Deep Think model earned a gold‑standard 42‑point score at the International Mathematical Olympiad (IMO) and performed comparably at the ICPC World Finals. In February 2026 three blog posts announced a major upgrade to the Deep Think inference mode, a new Gemini 3.1 Pro model, and positioned Deep Think as a "human intelligence multiplier".

The upgraded system delivered hard metrics: 48.4% on Humanity's Last Exam (no tool assistance), 84.6% on ARC‑AGI‑2 (officially verified by the ARC Prize Foundation), a Codeforces Elo of 3455, and gold‑level performance on the 2025 International Physics and Chemistry Olympiads. Detailed language‑by‑language results show perfect scores in Japanese (JMO Finals) and French, 86.3% on the Chinese Mathematics Olympiad (CMO) versus 63.3% on the Chinese Informatics Olympiad (NOI), and strong outcomes in Korean, Hindi, Vietnamese, Russian, and Portuguese.

The article questions the reliability of these numbers because the evaluation methodology is undisclosed: all scores come from internal Google testing, there is no third‑party replication, the number of inference runs per question, compute budget, or any prompt‑engineering assistance are omitted, and the contests referenced are regional qualifiers rather than international finals.

Google’s motivation for the eight‑language suite is explained as addressing the current bias toward English in AI benchmarks such as MATH, GSM8K, HumanEval, and ARC‑AGI. The selected languages cover major research hubs in East Asia, emerging markets, and Europe/South America, representing a large share of global scientific output. By demonstrating comparable performance across these languages, Deep Think aims to remove the language barrier for non‑English‑speaking researchers.

Beyond competition scores, DeepMind introduced Aletheia, a mathematics‑research agent powered by Deep Think that can autonomously generate, verify, and revise research‑level problem solutions. Aletheia has already contributed to papers, including one fully AI‑written study that computed a specific constant in arithmetic geometry, and solved four previously open problems from a set of 700 open mathematical questions.

In other domains, Deep Think has helped overturn a decade‑old conjecture in computer science, discovered a new analytic solution for cosmic string radiation in physics, and extended an auction‑theory theorem in economics. These achievements suggest the system is moving from pure competition performance to a broader "AI research accelerator" role.

The article concludes that while the scores themselves are impressive, the real signal is the engineering effort to treat multilingual AI reasoning as a solvable problem, potentially leveling the playing field for scientists worldwide. The next steps will involve independent verification and comparison with competing models.