By ByteBot
Google and OpenAI both launched major AI research products on December 11, 2025. This synchronized timing isn’t coincidental—it’s the clearest sign yet that the AI research race has entered a new phase of competitive escalation. Google made Gemini Deep Research available to developers for the first time via the Interactions API, while OpenAI countered with GPT-5.2. Moreover, the bigger story is OpenAI declared “code red” internally after Google’s Gemini 3 beat ChatGPT on benchmarks, reversing the roles from 2022 when Google panicked after ChatGPT launched.
What Gemini Deep Research Actually Delivers
Google’s Gemini Deep Research is now accessible to developers through the new Interactions API, a unified interface for working with Gemini models and agents. Powered by Gemini 3 Pro with a 1-million-token context window, the agent runs autonomously for hours or days, iteratively planning queries, identifying knowledge gaps, and searching again to fill them.
Furthermore, developers can access it via agent="deep-research-pro-preview-12-2025" through Google AI Studio. The Interactions API handles server-side state management, reducing boilerplate code for long-running research tasks. Use cases range from due diligence and pharmaceutical safety research to multi-step information synthesis with citation-rich reports.
Performance benchmarks show 46.4% success on Humanity’s Last Exam, 66.1% on Google’s new DeepSearchQA benchmark, and 59.2% on BrowseComp web browsing tasks. Notably, it trails GPT-5 Pro slightly on BrowseComp but leads on Google’s own benchmarks.
The Reality Check: Research Agents Still Need Human Oversight
Here’s the contradiction Google doesn’t emphasize: Gemini 3 Pro has an 88% hallucination rate—unchanged from Gemini 2.5 Pro. While the model improved at getting answers correct (53% accuracy, highest among current models), when it’s wrong, it’s overconfident. The Decoder notes that this hallucination rate means the model gives false answers rather than admitting uncertainty.
A 46.4% success rate on Humanity’s Last Exam means the “research agent” fails more than half the time. The “most factual model” claim sits uncomfortably alongside an 88% hallucination rate. Consequently, developers building with Deep Research still need human-in-the-loop verification and fact-checking—the agent branding sets expectations the technology doesn’t yet deliver.
OpenAI on Defense for the First Time
The same-day launch timing tells a deeper story. In early December, OpenAI CEO Sam Altman declared “code red” in an internal memo, redirecting resources to speed up ChatGPT improvements. Fortune reports this came after Google’s Gemini 3 launch in November put OpenAI on defense for the first time in years.
The stock market validated Google’s momentum. For the first time in nearly a decade, analysts say “Gemini/TPU stocks trade at a premium relative to ChatGPT/GPU peers.” Indeed, Gemini reached 650 million monthly active users and achieved a 1501 Elo score on LMArena—the first model to cross the 1500 threshold.
OpenAI fast-tracked GPT-5.2 to counter Google’s momentum, originally planning it for later in December. The company claims 70.9% expert-level performance versus Gemini 3 Pro’s 53.3%, though each company wins on their own benchmarks. This is a role reversal from 2022, when Google issued “code red” after ChatGPT disrupted search. Now OpenAI is shipping reactively.
What This Means for Developers
Competition benefits developers with more options and faster innovation. TechCrunch’s coverage of the simultaneous launches highlights the intensity of the AI race. Stack Overflow’s 2025 survey shows 78% of developers now use or plan to use AI tools, and 85% of enterprises plan to adopt AI agents by year’s end.
However, reactive shipping under competitive pressure creates risks. The Interactions API is in beta with features subject to breaking changes. Research agents from both companies still require oversight despite autonomous branding. Therefore, developers need to weigh stability against cutting-edge capabilities when choosing platforms.
The AI research wars are accelerating, with Google and OpenAI now shipping on identical schedules to signal competitive strength. Developers gain access to powerful autonomous research tools, but the 46.4% success rates and 88% hallucination rates mean fact-checking remains mandatory. Research agents are evolving rapidly—just don’t believe the marketing claims without testing the reality.
