Google Bets on Cheaper, Faster Gemini Models to Close the Gap on Rivals

Google is rolling out three new Gemini models on Tuesday, and the common thread running through all of them is price. As Alphabet heads into earnings with a product pipeline that's faced delays and increasingly aggressive competition, the company is leaning hard on cost and efficiency to make its case.

The most notable of the three is Gemini 3.5 Flash Cyber, a specialized model built to find and patch software vulnerabilities. It's Google's most direct move yet into a space where Anthropic has built a head start: automated code defense. For now, access is limited. The model is only available to governments and trusted partners through a pilot program, not the general public. Google says it runs at a lower price per token than its larger models, a detail that matters because cybersecurity work often means scanning enormous volumes of code, where cost per query adds up fast.

The other two releases target the broader developer and enterprise market. Gemini 3.6 Flash is a step up from the previous Flash model, with improvements to coding, multimodal understanding and general knowledge-work tasks. Google says it does this while using up to 17% fewer tokens and costing less per token than its predecessor. Tokens are the chunks of text AI models process to generate a response, and fewer tokens per task translates directly into lower bills for companies running these models at scale.

Then there's Gemini 3.5 Flash-Lite, which Google is positioning as its fastest and cheapest model in the 3.5 lineup. It's meant for high-volume, lightweight jobs, particularly the kind of small subtasks that make up larger AI-agent systems, where one main model might farm out dozens of simple requests to something leaner and quicker.

Racing on price, not just power

Google is framing the whole lineup around a simple pitch: even if it isn't always first with the most powerful model, it can win on efficiency. Data from Artificial Analysis backs that up to a degree. Gemini Flash already beats comparable models from Anthropic, OpenAI and Chinese developers on cost, according to the firm's figures. Google says Gemini 3.6 Flash, the stronger of Tuesday's two mainstream releases, costs less per task than GPT-5.6 Terra Max, Kimi K3 and Qwen 3.7 Max, while Flash-Lite comes in at a fraction of that price.

That pricing strategy lands at a pointed moment. Alphabet reports earnings the day after this rollout, and the AI landscape it's competing in has gotten more crowded and more urgent. Moonshot AI's Kimi K3 has seen demand strong enough that the company had to cap new subscriptions and API access because it couldn't keep up with capacity. Alibaba, meanwhile, is teasing Qwen 3.8 Max, which the company claims trails only Anthropic's Fable 5 in overall performance.

That kind of demand surge points to a problem that goes beyond who builds the smartest model. Serving an AI model to millions of users at scale requires enormous computing capacity, and that's turning into its own competitive battleground. Google may have an edge here: it designs its own custom chips, runs its own cloud infrastructure, and can build hardware and software in tandem rather than relying on outside suppliers. Still, the company has run into its own capacity limits before, so the advantage isn't automatic.

Custom silicon and what's next

Tuesday's launches arrive alongside reports that Google is developing a new specialized chip meant to run Gemini up to 10 times more efficiently, part of a wider effort to bring down the cost of serving AI at scale. 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," adding that "while not every project moves into production, this rigorous exploration is central to our full stack approach." The spokesperson also said that "by co-designing our hardware and software from the ground up, we ensure our systems are integrated and highly optimized for real-world workloads."

Google is also trying to get ahead of criticism that its roadmap has been murky, particularly after delays to prior releases. Gemini 3.5 Pro is now in testing with partners ahead of a wider rollout, and the company says it has kicked off its largest pre-training run yet for Gemini 4.

Put together, Tuesday's announcements read less like a single flashy launch and more like a company trying to prove it can compete on multiple fronts at once: specialized tools for high-stakes work like cybersecurity, cheaper everyday models for developers building at volume, and a hardware strategy meant to make all of it more affordable to run. Whether that's enough to close the gap with Anthropic and outpace fast-moving Chinese rivals will likely become clearer once Gemini 4 and Google's next-generation chips actually ship.