Google has added two new service tiers to the Gemini API that allow enterprise builders to manage the fee and reliability of AI inference relying on how time-sensitive a given workload is.
Whereas the price of coaching giant language fashions for synthetic intelligence has been a priority previously, the main focus of consideration is more and more transferring to inferencing, or the price of utilizing these fashions.
The brand new tiers, known as Flex Inference and Precedence Inference, tackle an issue that has grown extra acute as enterprises transfer past easy AI chatbots into advanced, multi-step agentic workflows, the corporate mentioned in a weblog submit revealed Thursday.
In a separate announcement on the identical day, Google additionally launched Gemma 4, the newest technology of its open mannequin household for builders preferring to run fashions regionally reasonably than by way of a paid API, describing it as its most succesful open launch so far.
The brand new API service tiers are meant to simplify life for builders of agentic programs involving background duties that don’t require on the spot responses and interactive, user-facing options the place reliability is crucial. Till now, supporting each workload varieties meant sustaining separate architectures: customary synchronous serving for real-time requests and the asynchronous Batch API for much less time-sensitive jobs.
“Flex and Precedence assist to bridge this hole,” the submit mentioned. “Now you can route background jobs to Flex and interactive jobs to Precedence, each utilizing customary synchronous endpoints.”
The 2 tiers function by a single synchronous interface, with precedence set by way of a service_tier parameter within the API request.
Decrease price vs larger availability
Flex Inference is priced at 50% of the usual Gemini API price, however affords decreased reliability and better latency. I is fitted to background CRM updates, large-scale analysis simulations, and agentic workflows “the place the mannequin ‘browses’ or ‘thinks’ within the background,” Google mentioned. It’s accessible to all paid-tier customers for GenerateContent and Interactions API requests.
For enterprise platform groups, the sensible worth is that background AI workloads equivalent to information enrichment, doc processing, and automatic reporting might be run at materially decrease price and not using a separate asynchronous structure, and with out the necessity to handle enter/output information or ballot for job completion.
Precedence Inference offers requests the best processing precedence on Google’s infrastructure, “even throughout peak load,” the submit acknowledged.
Nevertheless, as soon as a buyer’s visitors exceeds their Precedence allocation, overflow requests whereas not outright rejected are robotically routed to the Customary tier as a substitute.
“This retains your software on-line and helps to make sure enterprise continuity,” Google mentioned, including that the API response will point out which tier dealt with every request, giving builders visibility into each efficiency and billing. Precedence Inference is out there to Tier 2 and Tier 3 paid tasks.
However the downgrade mechanism raises issues for regulated industries, in accordance ot Greyhound Analysis Chief Analyst Sanchit Vir Gogia.
“Two an identical requests, submitted below totally different system circumstances, can expertise totally different latency, totally different prioritisation, and doubtlessly totally different outcomes,” he mentioned. “In isolation, this seems to be like a efficiency problem. In apply, it turns into an consequence integrity problem.”
For banking, insurance coverage, and healthcare, he mentioned, that variability raises direct questions round equity, explainability, and auditability. “Sleek degradation, with out full transparency and governance, is just not resilience,” Gogia mentioned. “It’s ambiguity launched into the system at scale.”
What it means for enterprise AI technique
The brand new tiers are a part of a broader business shift towards tiered inference pricing that Gogia mentioned displays constrained AI infrastructure reasonably than purely industrial innovation.
“Tiered inference pricing is the clearest sign but that AI compute is transitioning right into a utility mannequin,” he mentioned, “however with out the maturity, transparency, or standardisation that enterprises sometimes affiliate with utilities.” The underlying driver, he mentioned, is structural shortage — energy availability, specialised {hardware}, and information centre capability — and tiering is how suppliers are managing allocation below these constraints.
For CIOs and procurement groups, vendor contracts can now not stay generic, Gogia mentioned. “They need to explicitly outline service tiers, define downgrade circumstances, implement efficiency ensures, and set up mechanisms for price management and auditability.”
