Google has added two new service tiers to the Gemini API that enable enterprise developers to control the cost and reliability of AI inference depending on how time-sensitive a given workload is.
While the cost of training large language models for artificial intelligence has been a concern in the past, the focus of attention is increasingly moving to inferencing, or the cost of using those models.
The new tiers, called Flex Inference and Priority Inference, address a problem that has grown more acute as enterprises move beyond simple AI chatbots into complex, multi-step agentic workflows, the company said in a blog post published Thursday.
In a separate announcement on the same day, Google also released Gemma 4, the latest generation of its open model family for developers who prefer to run models locally rather than via a paid API, describing it as its most capable open release to date.
The new API service tiers are intended to simplify life for developers of agentic systems involving background tasks that do not require instant responses and interactive, user-facing features where reliability is critical. Until now, supporting both workload types meant maintaining separate architectures: standard synchronous serving for real-time requests and the asynchronous Batch API for less time-sensitive jobs.
“Flex and Priority help to bridge this gap,” the post said. “You can now route background jobs to Flex and interactive jobs to Priority, both using standard synchronous endpoints.”
The two tiers operate through a single synchronous interface, with priority set via a service_tier parameter in the API request.
Lower cost vs higher availability
Flex Inference is priced at 50% of the standard Gemini API rate, but offers reduced reliability and higher latency. I is suited for background CRM updates, large-scale research simulations, and agentic workflows “where the model ‘browses’ or ‘thinks’ in the background,” Google said. It is available to all paid-tier users for GenerateContent and Interactions API requests.
For enterprise platform teams, the practical value is that background AI workloads such as data enrichment, document processing, and automated reporting can be run at materially lower cost without a separate asynchronous architecture, and without the need to manage input/output files or poll for job completion.
Priority Inference gives requests the highest processing priority on Google’s infrastructure, “even during peak load,” the post stated.
However, once a customer’s traffic exceeds their Priority allocation, overflow requests while not outright rejected are automatically routed to the Standard tier instead.
“This keeps your application online and helps to ensure business continuity,” Google said, adding that the API response will indicate which tier handled each request, giving developers visibility into both performance and billing. Priority Inference is available to Tier 2 and Tier 3 paid projects.
But the downgrade mechanism raises concerns for regulated industries, according ot Greyhound Research Chief Analyst Sanchit Vir Gogia.
“Two identical requests, submitted under different system conditions, can experience different latency, different prioritisation, and potentially different outcomes,” he said. “In isolation, this looks like a performance issue. In practice, it becomes an outcome integrity issue.”
For banking, insurance, and healthcare, he said, that variability raises direct questions around fairness, explainability, and auditability. “Graceful degradation, without full transparency and governance, is not resilience,” Gogia said. “It is ambiguity introduced into the system at scale.”
What it means for enterprise AI strategy
The new tiers are part of a broader industry shift toward tiered inference pricing that Gogia said reflects constrained AI infrastructure rather than purely commercial innovation.
“Tiered inference pricing is the clearest signal yet that AI compute is transitioning into a utility model,” he said, “but without the maturity, transparency, or standardisation that enterprises typically associate with utilities.” The underlying driver, he said, is structural scarcity — power availability, specialised hardware, and data centre capacity — and tiering is how providers are managing allocation under those constraints.
For CIOs and procurement teams, vendor contracts can no longer remain generic, Gogia said. “They must explicitly define service tiers, outline downgrade conditions, enforce performance guarantees, and establish mechanisms for cost control and auditability.”
