The History of AI

Gemini 2.5 Pro

Era 5 · Generative AI, reasoning, and agents (2022–2026) › Alignment and reasoning

Lineage
Symbols and logic
Kind
Implementation

Who

Google DeepMind

What happened

Google DeepMind presents Gemini 2.5 Pro, the first of its Gemini 2.5 models, which it describes as thinking models: they reason before answering. It debuts at the top of LMArena with a million-token context window.

Why it mattered

Reasoning stops being a variant and becomes part of Google's main family of models.

The full story

After o1, reasoning had been treated as a separate line of models. Google had one too: Gemini 2.0 Flash Thinking, from December 2024. With Gemini 2.5 it changed course: every model in the new generation would be a thinking model, able to reason before answering.

The first was Gemini 2.5 Pro, presented on March 25, 2025, by Koray Kavukcuoglu, chief technology officer of Google DeepMind, as its most advanced model for complex tasks. It debuted at the top of LMArena, a leaderboard based on the preferences of users who compare two models' answers blind, by a margin Google called significant.

It also led several math and science benchmarks, such as GPQA and AIME 2025, and scored 18.8 percent on Humanity's Last Exam, a test of hard questions from many fields, the best figure among models not using tools. On SWE-bench Verified it reached 63.8 percent with a custom agent. Its context window, the amount of text it can take into account at once, was a million tokens, with two million announced.

With Gemini 2.5, reasoning became part of Google's main family, as it already had with Claude 3.7 Sonnet at Anthropic and Grok 3 at xAI.

Builds on

  • OpenAI o1 (September 12, 2024)

    Inherited, with no proven direct link

    After o1, the three labs add to their main models a mode that reasons before answering

Threads that pass through here

Sources

How to cite

“Gemini 2.5 Pro.” The History of AI, revised November 13, 2026. https://thehistoryofai.com/en/milestone/gemini-2-5-pro

This entry was last reviewed on November 13, 2026.

Open in the interactive timeline