Google, OpenAI and MiniMax Just Dropped Insanely Powerful AI at Once (Shocking Update)

OpenAI, Google, and MiniMax simultaneously launched powerful AI models, signaling a significant shift in software development through real-time coding, deep reasoning, and cost-effective autonomous agents.

• OpenAI's Codeex Spark offers real-time coding assistance through speed-optimized hardware and pipeline improvements.

• Google's Gemini 3 Deepthink provides advanced reasoning capabilities for complex scientific and engineering problems.

• MiniMax's M2.5 delivers highly cost-effective and autonomous AI agents for continuous professional tasks, emphasizing planning and integration into workflows.

OpenAI introduced GPT 5.3 Codeex Spark, optimized for real-time coding with near-instant inference, powered by Cerebras's specialized hardware. Google upgraded Gemini 3 Deepthink, focusing on advanced reasoning for scientific and engineering problems, featuring capabilities like sketch-to-3D creation. Concurrently, MiniMax released M2.5, a highly cost-effective agentic model designed for continuous, always-on operations across various professional tasks.

OpenAI's Codeex Spark

• 00:00:28 OpenAI launched Codeex Spark, a smaller, faster version of GPT 5.3 Codeex, specifically designed for real-time coding assistance where developers require near-instant responses to maintain flow. This model runs on Cerebras's wafer scale engine 3 (WSE3), a specialized hardware built for speed, indicating a strategic bet on mixed compute futures. OpenAI also optimized the entire request-response pipeline to reduce latency beyond just the model's speed, making interactions feel immediate for developers in Chat GPT Pro, CLI, and VS Code extensions.

Google's Gemini 3 Deepthink

• 00:05:41 Google unveiled Gemini 3 Deepthink, a specialized reasoning mode tailored for complex challenges in science, research, and engineering, focusing on problems with incomplete data and fuzzy constraints. Deepthink demonstrates strong performance across benchmarks like Humanity's Last Exam, ARC AGI 2, Code Forces ELO, and the International Math Olympiad. It utilizes 'test-time compute' to internally verify steps and prune bad paths, enhancing reliability, and also features a 'sketch-to-3D printing' demo, highlighting its capability to transform fuzzy human input into concrete artifacts.

MiniMax's M2.5 Agents

• 00:08:27 MiniMax introduced M2.5, an agentic model focused on economic practicality, making agents affordable to run continuously. Trained with reinforcement learning across vast real-world environments, M2.5 excels in coding, tool use, search, and office work, achieving high benchmarks on SWE bench and browse comp. The model is notably cost-efficient, allowing continuous operation for an hour at 100 tokens per second for about $1, and is designed to plan like an architect before coding, ensuring better-structured outputs and covering the full life cycle of development tasks across multiple languages and environments.

The Future of AI Integration

• 00:11:03 MiniMax's M2.5 agents emphasize the importance of robust tool calling and search as prerequisites for autonomous work, showcasing evaluations like browse comp and a new benchmark called RISE for realistic interactive search on professional tasks. The model claims to complete 30% of MiniMax's internal tasks autonomously and accounts for 80% of newly committed code, indicating deep integration into the company's operating system. This approach, powered by an in-house reinforcement learning framework 'Forge,' moves beyond language learning to enable AI to operate efficiently within real workflows and achieve real outcomes, posing the question of whether faster feedback or deeper reasoning will ultimately matter more.