AI's Power Plays: The Real Movers Behind Models, Money, and Market Shifts

Apple confirmed Siri will use Gemini in 2026. That’s not a rumor, it’s a deal — potentially the single most consequential distribution agreement in AI history. Last week, we talked about Anthropic’s massive compute deals redrawing the battlefield. Today, Google just secured the ultimate prize: its model inside every new iPhone. That was the capstone on a month of relentless releases. OpenAI updated its flagship for hundreds of millions of users on May fifth. GPT-5.5 Instant is reporting a fifty-two-point-five percent reduction in hallucinations compared to its predecessor, and its score on the AIME 2025 math benchmark jumped from sixty-five point four to eighty-one point two.

They also claim to have solved an 80-year-old math problem, which has been verified... but we've seen this pattern of overclaiming on proofs before. Meanwhile, Alibaba previewed Qwen 3.7-Max on May twentieth. It’s another closed-weight model with a one million token context window, and it’s already ranking seventh in math on the LM Arena leaderboards. They also unveiled a new AI chip, which is a clear signal that the hardware race in China is accelerating, not slowing down. And what about Anthropic, the subject of last week’s power moves?

This month, they shipped no new model. Instead, they went all-in on enterprise. They signed deals with PwC and KPMG to deploy Claude across hundreds of thousands of professionals. They took a two hundred million dollar grant from the Gates Foundation. Then they formed an entirely new AI services company with Blackstone, Hellman & Friedman, and Goldman Sachs. Their API volume is up seventeen times, year-on-year. They’re playing a different game. Finally, SpaceX filed for an IPO. The filing revealed what many suspected: the burn rate for its Grok AI model is significant, highlighting the brutal financial pressure of competing at this level.

Let’s go back to Google. The Apple deal is about distribution, but the product that just launched is about a fundamental capability shift. Gemini Omni Flash was released on May nineteenth. This is not just another text-to-video generator. This is a conversational, multi-turn video editing model. It integrates image, audio, video, and text inputs. You can tell it to change a character’s shirt, preserve scene consistency across multiple new shots, and it does it. Google’s official blog says it includes an imperceptible digital watermark and verification tools, a direct answer to governance concerns.

The Verge's hands-on review did note some visual artifacts in complex scenes, so it's not perfect. But the direction is clear. The era of clicking a toggle to activate a separate ‘thinking’ mode is over. Frontier systems are now natively blending deep, multi-step execution paths into the base model. But the most disruptive model of the month might not have come from Google or OpenAI. It came from a Chinese lab called DeepSeek. Their new model, DeepSeek V4-Pro, is open-weight. It has one-point-six trillion parameters, with forty-nine billion active at any one time.

And its performance is staggering. It leads all coding benchmarks, scoring ninety-three-point-five percent on LiveCodeBench. That surpasses both Claude at eighty-four-point-seven percent and the new GPT-5.5 at eighty-five-point-three. It has a one million token context window, it’s licensed under MIT, and it’s priced at a fraction of western models. Here’s the problem. The U.S. has accused DeepSeek of industrial-scale distillation—that is, training their models on the outputs of American AI labs. This shadows the entire achievement. It forces a difficult question: does the origin of the data invalidate the breakthrough?

The model is trained on Huawei Ascend chips, offering frontier-class AI at a seventy-five percent permanent discount. This isn't just a new competitor. It's a new paradigm, challenging everything from pricing to intellectual property. What moved this month wasn't just a set of version numbers. The very definition of a competitive model changed. It's no longer about a single benchmark score. Now, it's about distribution, enterprise integration, and the uncomfortable reality of state-level competition. The game just got much more complex.