Gemini 3 Deep Think Hits 84.6% on ARC-AGI-2 and 3455 Elo on Codeforces

Google's Gemini 3 Deep Think scored 84.6% on ARC-AGI-2, 48.4% on Humanity's Last Exam, and achieved a 3455 Elo rating on Codeforces competitive programming. These scores were independently verified by the ARC Prize Foundation in February 2026. Gemini 3.1 Pro is now in preview (March 2026), currently leading the Artificial Analysis Intelligence Index. For developers choosing a model for coding, reasoning, and complex analysis tasks, these numbers change the calculus.

What ARC-AGI-2 Actually Tests

ARC-AGI (Abstraction and Reasoning Corpus) was designed by François Chollet specifically to test general reasoning that cannot be solved by pattern matching or memorization. The original ARC-AGI was largely solved by frontier models in 2024-2025. ARC-AGI-2 is significantly harder — it presents novel visual reasoning puzzles that require genuine problem decomposition, not recall.

The scoring context: the average human score on ARC-AGI-2 is approximately 60%. GPT-4o and Claude 3.5 Sonnet scored below 20% on early evaluations. Gemini 3 Deep Think at 84.6% — verified by the ARC Prize Foundation, not just Google's internal evaluation — represents a meaningful discontinuity from previous frontier model performance on this benchmark.

The question Chollet designed ARC-AGI to answer is whether a model is doing genuine reasoning or statistical pattern completion. A score of 84.6% from a system that had no access to the test puzzles during training (ARC-AGI-2 puzzles are novel by construction) is evidence of reasoning capability that goes beyond lookup.

Humanity's Last Exam: 48.4% in Context

Humanity's Last Exam (HLE) is a benchmark assembled from questions that represent the hardest known problems across academic disciplines — the kind of questions that appear at the end of PhD qualifying exams or in the hardest competition mathematics. Questions are contributed by domain experts specifically to be outside what current AI systems can answer from training data.

48.4% on HLE means Gemini 3 Deep Think correctly answers nearly half of problems designed to be at the frontier of human expertise. For comparison: the human expert score on HLE questions from their own domain is around 90%; the score across all domains (where experts are not specialists) is lower. A score of 48.4% across all domains represents strong generalist performance at expert difficulty.

This does not mean Gemini 3 Deep Think is approaching human expert performance in the domains where it scored poorly — the benchmark is specifically designed to surface where models fail. The 51.6% it gets wrong includes the hardest problems in each domain. But the 48.4% it gets right is a substantial increase over what any previous model has achieved.

Codeforces 3455 Elo: What It Means for Developer Workflows

Codeforces is the competitive programming platform used by serious competitive programmers globally. The Elo rating system there maps to approximate skill levels: 1200 is beginner, 1900 is expert, 2400 is Grandmaster, and 3000+ is International Grandmaster — a tier occupied by fewer than 100 active human competitors worldwide.

3455 Elo places Gemini 3 Deep Think above the highest active human competitive programmers. This is not the same as saying it is better than top developers at real-world engineering tasks — competitive programming problems are well-defined, have verified solutions, and test algorithmic skill in isolation. But it is a strong signal for:

Algorithm and data structure implementation. If you need an efficient solution to a graph problem, a dynamic programming challenge, or a tree traversal, Gemini 3 Deep Think can likely produce correct, optimized code.

Debugging complex logical errors. Competitive programming skill correlates with the ability to identify logical edge cases — exactly the capability needed to debug subtle off-by-one errors, race conditions, and incorrect algorithm implementations.

Code review for correctness. A model at 3455 Codeforces Elo can evaluate whether an implementation is algorithmically correct, not just whether it passes basic tests.

Gemini 3.1 Pro Preview and the Current Leaderboard

Gemini 3.1 Pro (not Deep Think — a separate, more accessible model) entered preview in March 2026 and is currently leading the Artificial Analysis Intelligence Index, which aggregates performance across coding, reasoning, and instruction following benchmarks.

The model lineup as of late March 2026:

• Gemini 3.1 Pro — preview, leading Artificial Analysis Index, faster and more accessible than Deep Think
• Gemini 3 Deep Think — maximum reasoning mode, slower, best for hard problems requiring extended thinking
• Claude 3.7 Sonnet — strong on coding and reasoning, particularly with extended thinking mode
• GPT-4.5 — strong on instruction following and creative tasks, reasoning benchmarks below Gemini 3
• Llama 4 Scout/Maverick — Meta's open-weight frontier models, strong for self-hosted deployment

For most developer use cases — coding assistance, code review, document analysis, RAG — Gemini 3.1 Pro is the practical choice: faster response time, lower cost than Deep Think, and top benchmark scores. Deep Think is for tasks where you can afford to wait for extended reasoning: complex mathematical proofs, adversarial security analysis, hard algorithmic problems.

What the Benchmarks Do Not Tell You

ARC-AGI-2 and HLE measure specific reasoning capabilities in controlled conditions. They do not measure:

Context window reliability. How well does the model perform at the 128K or 1M token context boundary? Benchmark evaluations typically use short prompts. Real developer workflows often involve very long documents.

Instruction following consistency. Does the model reliably follow format instructions, output constraints, and system prompt requirements across 1,000 API calls? Benchmarks test single-shot performance. Production deployments need consistency.

Hallucination rate on domain-specific facts. ARC-AGI-2 tests abstract reasoning. A model that scores 84.6% on abstract puzzles may still hallucinate specific API method names, library versions, or regulatory details.

Latency and cost at production scale. Deep Think's extended reasoning mode is slow. For latency-sensitive applications — real-time coding assistants, chat interfaces — response time matters more than peak benchmark score.

Key Takeaways

• Gemini 3 Deep Think scored 84.6% on ARC-AGI-2 (independently verified by ARC Prize Foundation) — first model to substantially exceed average human performance (60%) on this benchmark
• 48.4% on Humanity's Last Exam — nearly half of PhD-level expert questions answered correctly across all domains
• 3455 Elo on Codeforces — above all active human competitive programmers, strong signal for algorithmic code generation and debugging
• Gemini 3.1 Pro in preview (March 2026), leading Artificial Analysis Intelligence Index — faster, more accessible version for daily developer use
• ARC-AGI-2 at 84.6% is evidence of genuine reasoning, not pattern matching — the benchmark is specifically designed to test novelty
• Benchmarks do not capture context window reliability, instruction following consistency, hallucination rate, or production latency — evaluate these separately for your use case