The real cost of AI in a business is not what you pay to buy access to a model. It is what you pay every time the model runs, multiplied by how often your workflows trigger it.
For most companies that have moved past testing into actual production, agentic AI tasks, where the model takes multiple steps, checks its own work, and calls external tools, consume anywhere from 5 to 30 times more processing than a simple chat exchange. One enterprise analysis of 2.4 billion real API calls found that organizations routing all work through premium-tier models paid $18.40 per million tokens, while those using a tiered approach paid $2.31. That gap is not marginal. It compounds fast at volume.
This is the problem Google is directly responding to today.
Gemini 3.6 Flash generates fewer words to reach the same answer. On a standard coding task, it uses 17% fewer output words than the previous version. In certain software development tests, that reduction reaches 65%. Because output costs more than input across every major AI provider, models that say less while doing the same job cost less per completed task. Design firm Figma and legal research platform Harvey have both adopted it: Figma for faster design iterations, Harvey and Hebbia for reading financial documents, extracting charts, and drafting reports.
The pricing: $1.50 per million words fed in, and $7.50 per million words produced. That output price is cheaper than the previous generation. Google also built computer-use capability directly into the model, meaning it can interact with software on a screen without needing a separate technical layer to translate between the AI and the application.
Gemini 3.5 Flash-Lite is the budget tier. At $0.30 input and $2.50 output per million words, it handles the tasks that do not require deep reasoning: scanning receipts, processing documents in bulk, routing searches. It runs at 350 output words per second, the fastest in Google's current lineup. Teams can use it as a worker that handles high-volume simple tasks, while reserving the more capable model for decisions that need more reasoning.
On the security side, Google is entering a market that Anthropic has moved early on. Anthropic's Mythos model, launched in April through a program called Project Glasswing with partners including Microsoft, Amazon, and Apple, found thousands of critical vulnerabilities across major operating systems and browsers. The original pricing ran at $25 per million words in and $125 per million out. By June, Anthropic released Mythos 5 at a reduced rate, still roughly double the cost of its premium general-purpose model.
Google's counter is Gemini 3.5 Flash Cyber: a smaller, cheaper model fine-tuned for finding and patching code vulnerabilities, running inside a system called CodeMender where multiple instances check each other's findings before producing a single report for a human to sign off on. In testing, it found 55 confirmed issues in a major JavaScript engine, compared to 47 for the standard Flash model. It also found 10 issues no other model found at all. But access is restricted to governments and vetted partners for now, citing the same concern Anthropic raised: a model this capable at finding vulnerabilities could just as easily be used to exploit them.
The broader context: enterprise AI bills are growing not because models are getting more expensive, but because companies are running them far more often than they originally planned. Uber burned through its entire 2026 AI budget in four months after rolling out AI coding tools to 5,000 engineers. Sam Altman acknowledged in June that cost concerns have become the second most common issue he hears from customers. Cheaper, more efficient models do reduce the per-task cost. But a cheaper model that gets deployed in more places, running more workflows, more often, often results in a larger total bill than the expensive model it replaced.
For any organization already running AI workflows or planning to: the number on the pricing page is not your cost. Your cost is that number multiplied by your volume, adjusted for how many steps your workflow takes per task. Model efficiency matters, but it matters most when you have already accounted for how many times the model will actually run.