Google adds cheaper Gemini models as AI cost race intensifies

Alphabet’s Google released three new Gemini models on Tuesday, including a cybersecurity-focused system and lower-cost Flash variants aimed at reducing the expense of running AI at scale. The launch came one day before Alphabet’s earnings report, with investors watching whether Google can improve the timing, capacity and economics of its AI products as rivals in the U.S. and China expand.

The most specialized release, Gemini 3.5 Flash Cyber, is designed to identify and repair software vulnerabilities, according to Google. The company said the model will first be offered through a limited-access pilot for governments and trusted partners, and that it runs at a lower price per token than larger models.

The product gives Google a clearer response to Anthropic’s early position in automated code defense, CNBC reported. Cybersecurity has become a prominent use case for frontier AI models because the systems can review code, detect flaws and propose fixes faster than manual review in some workflows, though access controls remain central for sensitive deployments.

Lower token use and cheaper tasks

Google is also introducing Gemini 3.6 Flash, which the company said improves performance on coding, multimodal tasks and knowledge work. The model uses up to 17% fewer tokens and costs less per token than its predecessor, according to Google.

Tokens are the units AI providers use to meter model input and output. For companies running large numbers of prompts, fewer tokens and lower per-token pricing can reduce the cost of each task, particularly in customer support, software development, document analysis and agent-based systems.

The third model, Gemini 3.5 Flash-Lite, is the fastest and least expensive model in Google’s 3.5 family, according to the company. Google said it is intended for high-volume workloads and smaller tasks inside larger AI-agent systems.

Artificial Analysis data cited by CNBC shows that Gemini Flash already costs less than comparable models from Anthropic, OpenAI and Chinese competitors. According to that data, Gemini 3.6 Flash is cheaper per task than GPT-5.6 Terra Max, Kimi K3 and Qwen 3.7 Max, while Gemini 3.5 Flash-Lite costs a fraction of that level.

Competitive pressure from China and Anthropic

The rollout comes as Chinese AI companies gain users and attention. Moonshot AI’s Kimi K3 attracted enough demand that the company limited new subscriptions and API access because of capacity constraints, according to a post from Kimi on X cited by CNBC. Alibaba has also been teasing Qwen 3.8 Max, which the company said trails only Anthropic’s Fable 5 in overall performance.

The demand pressures point to a broader constraint in the AI business: model quality alone does not determine commercial success. Providers also need enough computing capacity to serve customers reliably, especially when usage rises quickly after a model release.

Google may benefit from its custom chips, cloud infrastructure and ability to design models and hardware together, CNBC reported, though the company has also faced capacity constraints. CNBC reported this week that Google is developing a specialized chip intended to run Gemini up to 10 times more efficiently, part of a wider effort to reduce AI serving costs.

A Google Cloud spokesperson told CNBC that the company’s teams are “constantly researching and experimenting with new innovations to deliver maximum performance and efficiency for our users and customers.” The spokesperson added that Google’s approach is based on co-designing hardware and software so systems are optimized for real-world workloads.

Google is also giving more detail on its model roadmap after questions about delays, CNBC reported. Gemini 3.5 Pro is being tested with partners before a broader release, while the company has started its largest pre-training run to date for Gemini 4.

This story draws on original reporting from CNBC.