Google Gemini 3 Deep Think Hits 84.6% on ARC-AGI-2, Beating GPT-5 and Claude

Google released a major upgrade to Gemini 3 Deep Think on Thursday, its specialized reasoning model built for scientific and engineering work. The model posted an 84.6% score on ARC-AGI-2, a reasoning benchmark designed to resist memorization, according to results the ARC Prize Foundation verified independently.

That score beats OpenAI's GPT-5.2 Thinking at 52.9% and Anthropic's Claude Opus 4.6 Thinking at 68.8% on the same test. Humans average about 60%. The gap between Google's model and its closest rival, 15.8 percentage points, is nearly double the distance between that rival and average human performance.

Google said the update also reached 48.4% on Humanity's Last Exam without external tools, hit a 3,455 Elo on Codeforces competitive programming, and achieved gold-medal-level results on the 2025 International Olympiads in math, physics, and chemistry.

A benchmark built to resist AI is running out of room

Most benchmarks reward pattern recall from training data. ARC-AGI-2 does not. It tests whether a model can solve problems it has never encountered. It presents novel visual reasoning puzzles that require flexible, on-the-fly logic. Previous AI models struggled to break 20% on the original ARC-AGI. Gemini 3 Deep Think now scores 96% on that first version and 84.6% on the harder sequel.

The ARC Prize Foundation confirmed these numbers at a cost of $13.62 per task. That price matters. Deep Think's reasoning mode uses extended "test-time compute," spending more processing cycles before answering. Better scores come with higher inference bills. No enterprise customer has disclosed what that tradeoff looks like at production volume.

For comparison, Google's own Gemini 3 Pro Preview, the non-reasoning variant, scored 31.1% on ARC-AGI-2. The specialized reasoning mode is doing the heavy work.

The lead is consistent across every major test

The ARC-AGI-2 gap is not a one-benchmark anomaly. Deep Think scored 48.4% on Humanity's Last Exam without tools, beating Claude Opus 4.6 at 40% and GPT-5.2 at 34.5%. That benchmark throws thousands of PhD-level questions at models across specialized fields. Google's own Gemini 3 Pro Preview managed 37.5%.

In competitive programming, the margin is even wider. Deep Think holds a 3,455 Elo on Codeforces, "Legendary Grandmaster" territory that only a sliver of human programmers reach. Claude Opus 4.6 sits at 2,352. Anthropic and OpenAI should find that uncomfortable.

Multimodal reasoning and theoretical physics tell the same story. Deep Think scored 81.5% on MMMMU-Pro versus GPT-5.2 at 79.5% and Claude at 73.9%, and hit 50.5% on the CMT-Benchmark for advanced physics. Google is selling the model as a research partner, not a chatbot. These numbers make the pitch easier.

From benchmark charts to lab benches

Google paired the numbers with real-world demonstrations. Lisa Carbone, a mathematician at Rutgers University working on structures that connect general relativity and quantum mechanics, used Deep Think to review a technical paper. The model caught a subtle logical flaw that had passed through human peer review unnoticed, according to Google.

At Duke University, the Wang Lab used Deep Think to optimize fabrication methods for crystal growth. It designed a recipe for thin films larger than 100 micrometers, hitting a precision target that earlier methods had missed.

Google also introduced Aletheia, a math research agent powered by Deep Think that can run autonomous investigations or collaborate with humans. The agent can "admit failure to solve a problem," which Google said improved efficiency for researchers by avoiding dead-end paths. The company published several papers produced with the technology, from information theory to mechanism design.

Google demonstrated one more trick: feeding Deep Think a hand-drawn sketch and getting back a 3D-printable file. The model reads the geometry, writes code to recreate the shape, and spits out something a printer can actually use. Napkin to prototype, no CAD software in between.

Who gets access, and what it costs

If you are not a Google AI Ultra subscriber, you will not see this model for a while. The updated Deep Think is available now in the Gemini app for Ultra users. For the first time, Google is also offering it through the Gemini API, but access is limited to an early access program targeting researchers and enterprise users.

That restricted rollout reflects the model's inference costs. At $13.62 per ARC-AGI-2 task, running Deep Think at scale gets expensive fast.

Bloomberg reported the update as part of a broader push by AI labs to build tools for scientific research and complex coding. Anthropic recently released a new version of its most powerful model for financial research and legal analysis, triggering a selloff in traditional software stocks.

The benchmark saturation question hangs over all of these releases. ARC-AGI-1 is already at 96%. If its harder successor follows the same trajectory, the industry will need tougher tests to separate models, or shift the conversation to real-world output entirely.

Google's Gemini 3 Reveals What "PhD-Level Intelligence" Actually Means

Google launched Gemini 3 today with benchmark scores that dominate leaderboards and marketing copy promising "state-of-the-art reasoning" with "unprecedented depth and nuance." The model tops LMArena

The Implicator

DeepSeek's Math Model Beats Humans at Their Own Game, Then Gives Away the Playbook

Hangzhou-based AI lab DeepSeek dropped something unusual on November 27, 2025. Not another chatbot. Not incremental benchmark gains. A 685-billion-parameter math specialist that matches OpenAI and Goo

The Implicator

Alibaba's Qwen3-VL Can Find a Single Frame in Two Hours of Video. The Catch? It Still Can't Outthink GPT-5.

Alibaba just released a 42-page technical report detailing how its latest vision-language model processes two-hour videos with 99.5% accuracy on frame retrieval. The flagship Qwen3-VL-235B-A22B, built

The Implicator