Google DeepMind's Gemini 2.5 Pro arrived in early 2025 with benchmark numbers that turned heads across the AI industry. On the MMLU Pro, GPQA Diamond, and LiveCodeBench evaluations, it out-scores GPT-4o and matches or exceeds Claude 3.7 Sonnet on most tasks. The model's 1-million-token context window is the largest available in a production commercial API, enabling document analysis use cases that were previously impractical.
Yet benchmark leadership has not translated cleanly into enterprise market share. According to Andreessen Horowitz's 2025 AI survey of Fortune 500 procurement teams, OpenAI still accounts for over 60% of primary LLM API spend — a figure largely unchanged from 2024. Gemini's share sits at roughly 14%, followed by Anthropic's Claude at 18%.
Several factors explain the gap. First, OpenAI has an 18-month head start in enterprise sales infrastructure — dedicated account teams, SLAs, and compliance certifications (SOC 2 Type II, HIPAA BAAs) that procurement teams require. Second, the ChatGPT brand has become a de-facto interface standard; many enterprises built internal tooling around OpenAI's API patterns and switching carries real migration costs.
Google is attacking the gap on multiple fronts. Gemini is now deeply integrated into Google Workspace (Docs, Sheets, Meet), giving it a distribution advantage for the 3+ billion existing Workspace users. Vertex AI provides the enterprise compliance wrapper. And Google's ability to run inference at hyperscaler margins means it can undercut OpenAI on price — Gemini 2.5 Flash is currently priced at roughly one-fifth of GPT-4o for equivalent token volumes.
The real test is whether Google can convert its Workspace installed base into active Gemini API consumers. Early indicators from Q1 2025 earnings — Google Cloud grew 28% year-over-year, partly attributed to AI product revenue — suggest the strategy is gaining traction, even if OpenAI holds its top position for now.
For developers evaluating models, the practical advice is to benchmark on your own data rather than trusting leaderboard rankings — model quality varies significantly by task domain.