Google DeepMind has released Gemini Ultra 2, and the benchmark results are unambiguous: for the first time since GPT-4 launched in 2023, a non-OpenAI model holds the top position across every major evaluation suite. The release marks a genuine inflection point in the AI race — and has significant implications for enterprises, developers, and anyone who relies on large language models in their daily work.

The Benchmark Sweep

On MMLU (Massive Multitask Language Understanding), Gemini Ultra 2 scores 92.1%, compared to GPT-5's 90.8%. On MATH, the gap is even wider: 87.4% versus 84.2%. Most significantly, on the new GPQA Diamond benchmark — designed to test graduate-level scientific reasoning across physics, chemistry, and biology — Gemini Ultra 2 achieves 71.3%, a full 4 points ahead of its closest competitor.

These are not marginal improvements. GPQA Diamond is widely regarded as the most rigorous publicly available benchmark for scientific reasoning, and a 4-point lead represents a meaningful capability gap. Independent researchers at Stanford's CRFM (Center for Research on Foundation Models) who ran their own evaluations confirmed the results, noting that Gemini Ultra 2 showed particular strength in multi-step reasoning tasks that require integrating information across long contexts.

Multimodal Dominance

Where Gemini Ultra 2 truly separates itself is in multimodal tasks. The model can process video, audio, images, and text simultaneously with a 2-million-token context window — the largest of any publicly available model. In internal tests published in Google's technical report, it correctly answered detailed questions about a 3-hour documentary after watching it in real time, maintaining coherent understanding of narrative arc, character development, and factual claims throughout.

For practical applications, this means Gemini Ultra 2 can analyse an entire codebase, review a full legal document, or process a year's worth of financial reports in a single context window. The implications for enterprise use cases — due diligence, compliance review, research synthesis — are substantial.

The Architecture

Google has not disclosed full architectural details, but has confirmed that Gemini Ultra 2 uses a mixture-of-experts (MoE) approach with 1.8 trillion total parameters, of which roughly 280 billion are active on any given inference. This architecture — similar in principle to what GPT-4 is widely believed to use — allows the model to achieve high quality at significantly lower inference cost than a dense model of equivalent capability.

The MoE approach also enables more efficient specialisation. Different expert networks within the model appear to activate for different types of tasks — scientific reasoning, code generation, creative writing — allowing the model to apply specialised capacity where it is most needed rather than averaging across all tasks.

Pricing and Availability

Gemini Ultra 2 is available through Google AI Studio and the Gemini API. Pricing is set at $0.0125 per 1,000 input tokens and $0.0375 per 1,000 output tokens for the full model — roughly comparable to GPT-5 Turbo but with a significantly larger context window included at no additional cost. A Gemini Ultra 2 Flash variant, optimised for speed and cost, is priced at $0.00375 per 1,000 input tokens.

For developers already in the Google ecosystem — using Vertex AI, Google Cloud, or Workspace — integration is straightforward. For those on OpenAI's platform, migration requires API changes but the performance gains may justify the switch for latency-sensitive or cost-sensitive applications.

What This Means for the AI Race

The AI landscape has been dominated by OpenAI's narrative for three years. GPT-4 set the benchmark in 2023; GPT-4o refined it in 2024; GPT-5 extended it in 2025. Gemini Ultra 2 doesn't just challenge that narrative — it inverts it. For the first time, a Google model is the clear leader across the benchmarks that matter most to enterprise buyers.

The competitive implications extend beyond benchmarks. Google's distribution advantages — integration with Search, Workspace, Android, and YouTube — give Gemini Ultra 2 a deployment surface that OpenAI cannot match through API access alone. If the model quality advantage holds, Google is positioned to convert its distribution scale into AI market share in a way that has not been possible with previous Gemini generations.

For enterprises evaluating AI infrastructure, the calculus has fundamentally changed. The question is no longer whether to use OpenAI — it is whether OpenAI's ecosystem advantages outweigh Gemini Ultra 2's capability and cost advantages. That is a much harder question than it was six months ago.

Sources & Further Reading