Google's upgrade breaks reasoning barriers

2026-02-13 · Source: The Rundown AI · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering, Emerging Technologies & Innovation · Depth: Intermediate, medium

Summary

Google has significantly upgraded its Gemini 3 Deep Think reasoning mode, achieving dominant scores across various benchmarks in math, coding, and science. The updated model scored 84.6% on ARC-AGI-2, surpassing Opus 4.6 (68.8%) and GPT-5.2 (52.9%), and reached 48.4% on Humanity's Last Exam. Deep Think also earned gold-medal marks on the 2025 Physics & Chemistry Olympiads and an Elo score of 3,455 on Codeforces. Concurrently, Google unveiled Aletheia, a math agent powered by Deep Think, capable of autonomously solving open problems and verifying proofs. This upgrade is available to Google AI Ultra subscribers in the Gemini app, with API access for researchers via an early access program. Separately, OpenAI released GPT-5.3-Codex-Spark, a speed-optimized coding model running on Cerebras hardware, achieving over 1,000 tokens per second, and MiniMax launched its open-source M2.5 model, which rivals frontier coding models at a significantly lower cost.

Key takeaway

For AI Architects and Machine Learning Engineers evaluating frontier models, Google's Deep Think and Aletheia demonstrate significant advancements in reasoning and autonomous problem-solving, suggesting a shift in capabilities for scientific AI. Your teams should explore the Deep Think early access program for research applications and consider MiniMax's M2.5 for cost-effective, high-performance coding agents to optimize operational expenses.

Key insights

Google's Deep Think and Aletheia agents are pushing AI reasoning and scientific discovery into uncharted territory.

Principles

• Speed can be prioritized over raw intelligence for specific AI tasks.
• Cost-effective frontier models can democratize advanced AI capabilities.

Method

To generate a TV commercial with AI, plan scenes and prompts with Gemini, then use Higgsfield's Image and Video tools (Nano Banana Pro, Kling 3.0) to create frames and stitch videos.

In practice

• Utilize Deep Think for advanced math and science research tasks.
• Deploy GPT-5.3-Codex-Spark for real-time coding edits.
• Consider MiniMax M2.5 for cost-efficient, high-performance coding agents.

Topics

• Google Deep Think
• OpenAI Codex-Spark
• AI Reasoning Benchmarks
• Open-source LLMs
• AI Agentic Systems

Best for: AI Architect, Machine Learning Engineer, AI Scientist, AI Engineer, Data Scientist, General Interest

Related on AIssential

• 🗞️ Google releases Gemini 3 Deep Think, tops ARC-AGI 2 Benchmark With 84.6% — Advanced AI models are achieving human-level performance in complex reasoning and code generation through enhanced architectural and operational methods.
• [AINews] OpenAI GPT-next disproves 80 year old Erdős planar unit distance problem for under $1000 — General-purpose LLMs are achieving significant scientific breakthroughs and driving architectural innovation.
• 2025 Interconnects year in review — DeepSeek's open-weight, efficient MoE and MLA architectures challenge AI's cost curve, intensifying the global AI race and geopolitical tech competition.
• The Secret Tech Google and OpenAI Use to Beat Anthropic — Google and OpenAI are heavily investing an estimated "quarter-trillion dollars" in "inference-time compute scaling" techniques for their latest Large Language Models, such as GPT-5.4 Thinking and Gemini 3.1 Pro.
• Am I completely insane for thinking AI is mid — Current LLMs excel at text prediction but fundamentally lack robust reasoning and intuition, leading to limitations and "hallucinations."
• Last Week in AI #335 - Opus 4.6, Codex 5.3, Gemini 3 Deep Think, GLM 5, Seedance 2.0 — Frontier AI models are rapidly advancing in multi-agent collaboration, extended context, and specialized reasoning across diverse domains.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by The Rundown AI.