Did AI Just Get Commoditized? Gemini 2.5, New DeepSeek V3, & Microsoft vs OpenAI

Gemini 2.5 Pro and DeepSeek V3 arrive with a clear message for the AI market: top-tier language-model performance is converging across companies, making “secret sauce” less visible than the amount of compute and engineering effort behind each system. The practical takeaway is that model leadership is becoming harder to sustain, even as models keep improving—because multiple labs are moving toward similar capability levels on the same kinds of tasks.

Google’s Gemini 2.5 Pro is positioned as its “most intelligent” model, and the transcript leans on a mix of benchmark categories to argue that it’s not just another incremental release. In knowledge-heavy testing—described as “humanity’s last exam,” with obscure trivia, Latin translations, and specialized science—Gemini 2.5 Pro is portrayed as leading for “knowledge without searching the web.” The comparison also highlights how OpenAI’s O3 mini (not the full O3) is expected to score higher once the larger model is released, and it notes that some benchmark reporting differences (like majority voting and tool use) can make direct comparisons tricky.

The same convergence theme shows up in science and math. Gemini 2.5 Pro is described as roughly level with OpenAI’s o3-mini and close to other frontier models when extended reasoning is allowed. The transcript also flags a key structural issue in benchmarking: different labs use different evaluation methods, including whether they apply majority voting or allow extra compute, which can widen or narrow gaps depending on the rules.

Where Gemini 2.5 Pro stands out most sharply is multimodal and long-context capability. It’s described as state-of-the-art at reading tables and charts in the MMU benchmark, and the transcript claims it’s the first model to get “within touching distance” of human performance there—while humans are said to have the advantage of browsing and taking their time. On long context, Gemini 2.5 Pro is said to handle up to a million tokens, far beyond other models in the cited chart.

DeepSeek V3 reinforces the same story from a different angle. The transcript distinguishes DeepSeek V3 as a new base model (not the separate reasoning model R2), and frames it as comparable to OpenAI’s GPT 4.5 base layer for future reasoning systems. In the benchmark comparisons, DeepSeek V3 is portrayed as notably stronger in mathematics and competitive in coding, while being closer to OpenAI on science and general knowledge. The broader claim: if base models are converging, then reasoning models also lack a clear moat.

That leads into the “commoditization” argument tied to Microsoft’s leadership. The transcript references Microsoft CEO claims that models are being commoditized—performance increasingly bought like a commodity through compute and scaling, with labs selling an experience rather than a unique path to AGI. A Microsoft internal unit (Microsoft AI) is described as having reverse-engineered or replicated “reasoning-like” behavior (the “think before answering” pattern associated with systems such as Gemini and DeepSeek’s R1), with Microsoft claiming near-parity on benchmarks with leading OpenAI and Anthropic models.

Finally, the transcript adds a reality-check on the hype cycle: even as predictions grow that AI will write nearly all code, the transcript points to ongoing hiring and a playful example of a model getting stuck in a simple game scenario. The overall conclusion: Gemini 2.5 Pro looks impressive, but it reads less like a singular breakthrough and more like evidence that frontier capabilities are converging across model families—making differentiation increasingly about deployment, cost, and product execution rather than undisclosed intelligence secrets.