AI Daily: Google Reasoning Evolution, MiniMax vs. OpenAI Speed War, Anthropic Valuation Skyrockets

It has been a wild weekend, with AI news flooding in like an avalanche. If you thought the previous pace of model updates was fast, the developments over the past two days might redefine your definition of “efficiency.” Today, we’re skipping the vague concepts and diving straight into the substance these four giants have delivered.

From Google enabling AI to think like a scientist, to the head-to-step confrontation between MiniMax and OpenAI in coding speed, and finally to Anthropic’s staggering valuation, every update points to the same trend: AI is no longer just a toy for chatting; it is becoming a practical tool for solving complex scientific problems and engineering challenges.

Google Gemini 3 Deep Think: Beyond Coding, It’s Doing Science

Remember the performance of Google’s previous models in math competitions? This time, they aren’t just aiming for medals; they truly want AI to solve real-world scientific problems. Google has just announced a major update to Gemini 3 Deep Think, a mode specifically born for reasoning.

Honestly, this upgrade is a bit spine-tingling. It no longer just processes data; it has learned to “think” like a scientist.

Reasoning Like a Nobel Laureate

Gemini 3 Deep Think’s performance in the scientific field is nothing short of dominant. It achieved gold-medal-level results in the written portions of the 2025 International Physics Olympiad and Chemistry Olympiad. Even more impressively, it scored 48.4% on “Humanity’s Last Exam,” a benchmark designed to test the absolute limits of models, without using any external tools.

Here’s a great example: Lisa Carbone, a mathematician at Rutgers University, used Deep Think to review a technical paper on high-energy physics. The AI managed to find a logical loophole that human peer reviewers had missed. This demonstrates its potential to assist top-tier researchers.

From Sketch to 3D Print

Beyond abstract theory, it has also become smarter in engineering applications. You can now draw a rough sketch on paper, and Deep Think can analyze the graphic, build a complex geometric model, and even directly generate a file ready for 3D printing. This is a massive boon for engineers who have ideas but find CAD drawing tedious.

MiniMax M2.5: The “Virtual Architect” Is Cheaper Than You Think

While Google pursues scientific extremes, MiniMax is chasing extreme productivity and cost-effectiveness. They have just released the MiniMax M2.5 model, with a core philosophy that is crystal clear: built for real-world productivity.

Specifications First, Code Later

This is perhaps the most interesting aspect of M2.5. While many AI models write code as they go, M2.5 demonstrates the qualities of an “architect” during its training. Before typing a single line of code, it acts like an experienced software architect, clearly planning out functionality, structure, and UI design.

This “think before you act” strategy helped it achieve a high score of 80.2% on SWE-Bench Verified (a software engineering benchmark). This means it’s not just fixing bugs; it can handle the complete development cycle from system design to feature iteration. Developers can check out its capabilities on HuggingFace.

Incredibly Fast, Practically Free

Here’s a crazy stat: M2.5’s inference speed reaches 100 tokens per second, nearly double that of other frontier models. Even more shocking is the price: running it at this speed continuously for an hour costs only $1. At 50 tokens per second, the cost drops to $0.30. This means we are one step closer to a future where “intelligence is too cheap to meter.”

OpenAI GPT-5.3-Codex-Spark: A High-Speed Marriage with Hardware

OpenAI hasn’t been idle either, clearly recognizing the importance of “speed” in real-time collaboration. They have introduced GPT-5.3-Codex-Spark. The “Spark” suffix signifies that this is an ultra-fast model specifically designed for real-time coding.

A Victory for Software-Hardware Integration

The most noteworthy part of this release isn’t the model itself, but OpenAI’s partnership with chip startup Cerebras. Codex-Spark runs on the Cerebras Wafer Scale Engine 3, hardware specifically designed for AI inference.

The result? A generation speed exceeding 1,000 tokens per second. Yes, 1,000. This makes the coding experience feel almost “instantaneous.” It solves a major pain point: when you’re coding, you don’t want to wait for the AI to spin; you want it to keep up with your thoughts.

Fine-Tuned for the “Now”

Unlike models specialized for long-range reasoning, Codex-Spark is designed as a lightweight assistant. It’s perfect for targeted edits, refactoring logic, or adjusting interfaces. While it currently only supports text and a 128k context window, this low-latency experience is exactly what developers need for rapid iteration.

Anthropic: The King of Enterprise AI

Finally, we must talk about the money. While Anthropic didn’t release a new model this time, they announced news that shocked the industry: the completion of a $30 billion Series G funding round.

Behind the $380 Billion Valuation

This round brings Anthropic’s valuation to a staggering $380 billion. What is this money for? More powerful computing power and infrastructure, of course. The investor list, led by GIC and Coatue and including Microsoft and NVIDIA, is a star-studded lineup. Furthermore, their models are now fully integrated into Amazon and Google’s cloud platforms.

This reflects a reality: the corporate world trusts Claude. Anthropic’s annual run-rate revenue has reached $14 billion, with more than 10x growth annually over the past three years. In particular, agents like Claude Code, which can autonomously complete coding tasks, are being adopted by more and more companies. This isn’t just funding; it’s a vote of confidence from the market in “safe and powerful enterprise AI.”

Frequently Asked Questions (FAQ)

Q: For a typical programmer, is MiniMax M2.5 or OpenAI Codex-Spark better?

It depends on your use case. If you need AI to help you plan an entire system from scratch or handle complex architectural issues that require deep thinking, MiniMax M2.5’s “architectural mindset” and extremely low cost might be a better fit—it excels at breaking down large tasks. However, if you are in the middle of writing code and need an assistant that can keep up with your typing speed to provide instant completions or small-scale refactors, OpenAI Codex-Spark’s 1,000 tokens per second will feel much smoother and won’t break your flow.

Q: Is Google’s Deep Think mode available now?

Yes, the new Deep Think mode is now available in the Gemini App for Google AI Ultra subscribers. For developers and enterprise users, Google has also opened early access to Deep Think via the Gemini API for the first time. However, keep in mind that this mode is specifically designed for deep reasoning; using it for everyday casual chat might be overkill.

Q: How does Anthropic’s massive funding affect average users?

In the short term, it means Anthropic has the capital to purchase more GPUs (like AWS Trainium and Google TPUs) to train next-generation models (such as the rumored Claude 4 or 5). This ensures the Claude series will remain among the most powerful competitors in the market for years to come, especially in handling long text and complex logic. It also suggests they will continue to strengthen Claude’s enterprise features, such as enhanced security and privacy protections.