Google Cuts Free Gemini API Access: Developers React

Google has significantly curtailed the daily request limit for its free Gemini API, reducing it from 250 to 20 requests. This change has rendered automation scripts and small development projects largely unusable for some developers, according to user Nilvarcus. The company also removed the Pro series from the free tier and limited the Flash series to 20 daily requests.

Other users noted that Google has removed the free Gemini API entry from its "Bulk API Rate Limits" list. This move follows a period where Google offered extensive free access to its Gemini API, including the Gemini 1.5 Flash free tier, which provided up to 1.5 billion free tokens daily, 15 requests per minute, and 1,500 requests per day, along with free context caching and fine-tuning features.

Developer Reactions to Unannounced Changes

The policy change was implemented without prior notification, which angered some developers. One developer stated that while they understood the concept of "no free lunch," Google's abrupt approach was unacceptable. They suggested that a responsible company would have announced such changes in advance, perhaps when launching Gemini 3.

Another developer commented that Google had likely gathered sufficient data and gained a competitive edge, leading to a strategic shift toward profitability. They suggested that the initial generosity of the free tier served to attract users and train models, and the company is now transitioning to paid conversions.

Competitive Landscape and OpenAI's Response

Google's Gemini 3 had recently gained traction, with user engagement data from the Financial Times indicating an average usage time of approximately 7.2 minutes on desktop and mobile web, surpassing ChatGPT's 6 minutes and Anthropic Claude's 6 minutes.

However, the competition in large language models remains intense. OpenAI is reportedly preparing to release GPT-5.2, initially planned for late December but now anticipated around December 9. Leaked benchmark results for GPT-5.2 suggest that it could restore OpenAI's competitive advantage.

Amid these rumors, Gemini 3 Flash became available on LM Arena, with some observers suggesting it is Google's direct response to GPT-5.2. Demis Hassabis, co-founder and CEO of Google DeepMind, emphasized Google's determination to maintain a leading position in AI, regardless of market fluctuations.

Google's Satisfaction with Gemini 3

Hassabis expressed satisfaction with Gemini 3's performance, highlighting its concise answers and ability to challenge unreasonable prompts. He noted that users perceive a significant advancement in intelligence, making the model more useful. This approach contrasts with OpenAI's previous experience with GPT-4o, which faced criticism for being overly compliant.

Hassabis also noted the rapid pace of user innovation with new AI technologies, stating that users often discover novel applications that internal teams have not yet explored. He cited Gemini 3's ability to create games and assist with front-end development as examples of its capabilities.

Pushing the Limits of Scaling Law

The sudden restrictions on the free Gemini API led some users to speculate about computational resource constraints. Hassabis acknowledged that while Google and DeepMind possess substantial resources, they are not infinite and consistently require more computing power.

He remains a proponent of the Scaling Law, believing that pushing current system scales to their limits is crucial for achieving Artificial General Intelligence (AGI). Hassabis estimates that AGI could be achieved within five to ten years, defining it as possessing all human cognitive abilities, including creativity and inventiveness.

Hassabis described current large language models as having "jagged intelligence," excelling in some areas but lacking consistency, continuous learning, long-term planning, and complex reasoning. He anticipates that one or two major breakthroughs will be necessary to address these limitations. He recounted Google's pragmatic approach to AGI research, prioritizing methods that demonstrate empirical effectiveness.

The Scientific Method as a Core Advantage

Hassabis views the scientific method as fundamental to Google's approach, emphasizing rigor, precision, and evidence-driven development. He stated that combining top-tier research, engineering, and infrastructure provides a competitive edge in the AI frontier.

Regarding AI talent, Hassabis acknowledged the intense competition but highlighted Google's focus on "mission-driven" individuals. He believes DeepMind's mission and comprehensive capabilities attract top scientists and engineers seeking impactful work.

Google's Future Directions in AI

Google is concentrating on three primary areas for AI development:

Modality Fusion: Gemini was designed as a multimodal model, processing and generating content across images, video, text, and audio. Hassabis noted the synergistic effects across modalities, citing the Nano Banana Pro image model's visual understanding and infographic generation capabilities. He anticipates significant advancements in video and language model fusion in the coming year. Hassabis highlighted the multimodal understanding of video, images, and audio, particularly in processing YouTube videos. He described an instance where Gemini provided a philosophical interpretation of a scene from "Fight Club," demonstrating its deep conceptual understanding. He also mentioned Gemini Live, a feature that allows users to interact with objects through their phone camera, envisioning its future integration into devices like smart glasses.

World Models: Hassabis is personally driving the development of world models, exemplified by Genie 3, an interactive video model. This system allows users to generate and then "enter" a video, maintaining coherence for approximately one minute, akin to a simulated world.

Agent Systems: Hassabis expects significant progress in agent systems, aiming for a "universal assistant" that integrates into various devices beyond computers and phones. He believes these agents will enhance productivity and personal life by providing recommendations and assistance. However, he acknowledged that current agents lack the reliability for full task delegation. Hassabis also noted the challenge of ensuring that increasingly autonomous agents remain within defined guardrails, emphasizing the need for ongoing research in this area. He suggested that commercial incentives, such as enterprise demands for reliability and data guarantees, will encourage responsible AI development.

Hassabis also commented on the global AI landscape, stating that while the U.S. and the West currently hold a lead, China is rapidly catching up, with strong models and capable teams. He estimated the lead to be "months" rather than "years," attributing the Western advantage in AI algorithm innovation to a tendency to propose new algorithms beyond current frontiers.