Google Positions Itself to Claim 2024 AI Crown wiht Gemini 3 Launch
Seoul, South Korea – November 20, 2024 – Google is making a strong bid for “AI trophy” of the year with the release of Gemini 3, showcasing significant performance gains alongside maintained usability and pricing. The new model, particularly in its “Deep Think” advanced reasoning mode, is demonstrating capabilities exceeding those of competitors, according to a report from IT Donga.
Gemini 3 Pro achieved 41% accuracy on “humanity’s final problem” in Deep Think mode, a significant enhancement over the 37.5% achieved in normal mode. The model also set a new record of 45.1% on the ARC-AGI 2 puzzle test,a benchmark requiring human-level flexible thinking – dramatically surpassing the 10-15% performance of existing models.
Google is releasing Gemini 3 in tiered subscription levels.’Google AI Pro’, including standard Gemini 3 access, remains priced at 29,000 won per month.’Google AI Ultra’, offering access to advanced users and Gemini 3 Deep Think, is priced at 360,000 won per month, excluding promotional offers.
While subscription costs remain stable, API integration costs for developers have increased. Gemini 2.5 Pro’s API previously cost $1.25 (approximately KRW 1,836) for input and $10 (approximately KRW 14,690) for output per 1 million tokens. The Gemini 3 Pro Preview API now costs $2 (approximately KRW 2,938) for input and $12 (KRW 17,632) for output per 1 million tokens - a 60% increase for input and a 20% increase for output. Though, Google is now offering control over costs through adjustable inference intensity settings (Low and High), replacing previous token limitations.
Industry reaction has been positive. While OpenAI’s GPT-5 underwent rapid iteration following its release, Gemini 3 Pro is already receiving praise from developers and users, with no reported major inconveniences. IT donga reports that expectations are high for future derivative versions like Flash or Nano, and potential performance improvements in other Google AI products such as Veo, Imagen, and Med-Gemini.
Reporter Nam Si-hyeon of IT Donga (sh@itdonga.com) contributed to this report.
