Google Cuts Agent Operating Costs and Separates Models by Speed and Risk

Google divided its new Gemini portfolio around three operating requirements: the quality of a general-purpose agent, the lowest possible price at high volume, and specialised cyber defence. The company introduced Gemini 3.6 Flash, Gemini 3.5 Flash-Lite and Gemini 3.5 Flash Cyber on July 21. Enterprises can therefore combine several models in a single service instead of using the most expensive option for every task. According to Google, Gemini 3.6 Flash uses 17 percent fewer output tokens than Gemini 3.5 Flash in the Artificial Analysis composite evaluation. In the DeepSWE coding-agent benchmark, savings reached as much as 65 percent on some tasks. The model also used fewer reasoning steps and fewer calls to external tools. Pricing is set at 1.50 dollars per million input tokens and 7.50 dollars per million output tokens. For companies, the decisive figure is the cost of a completed task rather than the price of a token alone. If the model uses fewer steps, spending on databases, search, code execution and review of intermediate results also falls. Gemini 3.6 Flash scored 49 percent in DeepSWE, compared with 37 percent for version 3.5 Flash. In MLE Bench, the score rose from 49.7 to 63.9 percent. In OSWorld-Verified, which tests work with graphical interfaces, the model scored 83 percent, up from 78.4 percent. Flash-Lite targets translation, receipt processing, search and high-volume document analysis. Google reports a speed of 350 output tokens per second. Pricing is 0.30 dollars per million input tokens and 2.50 dollars per million output tokens. In Terminal-Bench 2.1, the model scored 54 percent, compared with 31 percent for the previous Flash-Lite. Such a model can handle the first stage of a business process: classify documents, extract data or prepare a draft response. A more expensive model then receives only complex cases. This division can reduce costs, but it requires precise escalation rules. Incorrect classification may cause a cheaper model to process a legally or financially sensitive decision. Gemini 3.5 Flash Cyber has the greatest regulatory significance. Google tuned it for finding, validating and fixing vulnerabilities. Several agents in the CodeMender system analyse the same problem and produce a combined report. Because of potential misuse, Google will not release the model publicly. A limited pilot will be offered to governments and vetted partners. Google is therefore acknowledging the technology's dual-use nature. The same system can help secure software or show an attacker how to break it. Restricted access reduces availability but does not resolve accountability. Organisations must record who used the model, which software was tested and what happened after a flaw was identified. “Developers need greater token efficiency, lower latency and more reliable performance,” said Tulsee Doshi, product director for the Gemini team. Her wording captures the commercial shift: competition is moving away from model size toward the cost of work that is actually completed. Google also announced that it had begun its most demanding initial training run yet for Gemini 4. That means further capital expenditure on computing infrastructure before investment in the current generation has paid back. Customers should therefore address access to older versions, pricing changes and the impact of automatic updates on validated processes in their contracts.

Google Cuts Agent Operating Costs and Separates Models by Speed and Risk

The full article is available to users registered at Hard Skills