What Oracle added
Gemini is the newest name in Oracle's third-party model catalogue, and it is not confined to the infrastructure layer. The company is putting it into OCI AI Agent Studio, into Fusion Cloud Applications and into NetSuite, so a finance or HR team can call the model inside the application it already runs, without signing a separate Google Cloud agreement. The expansion was set out in the joint Oracle and Google Cloud announcement of 30 July 2026.
The move follows the pattern Oracle has used for two years: sell infrastructure, then sell access to every credible frontier model on top of it. Meta's models, OpenAI's models, xAI's Grok, Cohere and Alibaba's Qwen are already on that list. Oracle does not need its own frontier model to win enterprise AI deals, and it is not building one.
Wall Street reacted to the distribution story rather than to the model itself. Neither company has published revenue figures for Gemini routed through OCI. Oracle's capital spending on AI capacity, not the share move, is the more honest measure of how much hardware the company expects to sell.
One contract, six model vendors
A European enterprise that already holds an Oracle agreement can add a model without a new vendor onboarding, a new data processing agreement or a fresh security review. In large organisations those steps frequently take longer than the integration work itself.
The setup also creates a switching option. If a model's price rises, if latency degrades, or if a provider lands on the wrong side of a compliance question, traffic can be routed to a different model behind the same API pattern and the same logging layer. That is the argument Oracle is selling, and it is a reasonable one.
Prices per million tokens
Published Google list price for the Gemini model confirmed in Oracle's catalogue, per million tokens:
| Model | Input (USD / 1M) | Output (USD / 1M) |
|---|---|---|
| Gemini 3.8 Flash (introductory, through 31 Dec 2026) | 0.75 | 3.75 |
| Gemini 3.8 Flash (from 1 Jan 2027) | 1.50 | 7.50 |
Oracle does not publish one blended rate for the whole catalogue, and negotiated enterprise pricing rarely matches list. Other Gemini variants — names such as 3.5 Flash and 3.1 Flash-Lite appear in circulating catalogue references — could not be confirmed against both vendors' published lists, so they are not priced here. Treat the figures above as a starting point for the conversation, not as an OCI price list.
What a European customer pays on top
Two costs sit between the list price and the invoice.
VAT comes first. OCI list prices are quoted in US dollars, and VAT is added according to the billing country. Standard rates run from 17% in Luxembourg to 27% in Hungary, with 21% in Czechia, 19% in Germany and 23% in Ireland. Take Gemini 3.8 Flash output tokens: at the introductory $3.75 per million, a 50 million-token month costs $187.50 at list and $226.88 on a Czech invoice with 21% VAT. From 1 January 2027 the same model's output price rises to $7.50 per million, which doubles that workload to $375 at list and $453.75 with Czech VAT — and a 100 million-token month reaches $750 at the 2027 list price and $907.50 on the invoice.
Currency comes second. Because the invoice is denominated in dollars, the euro cost moves with the exchange rate, and finance departments budgeting in EUR carry that exposure on every AI line item. Some Oracle EU regions support billing in local currency; that is worth confirming before committing to a multi-year figure.
What the AI Act requires
Buying a model through a cloud marketplace does not remove obligations under the EU AI Act. General-purpose AI obligations have applied since August 2025, and transparency and deepfake-labelling duties became enforceable on 2 August 2026. The Digital Omnibus amendment pushed full compliance for high-risk systems to 2027 and 2028, but the prohibited-use rules and the transparency rules stayed in place.
In practice, the model provider carries the GPAI documentation duties. The company deploying the model still has to document its own use case, keep logs where required, and be able to explain which model produced which output. A catalogue with six vendors makes that harder unless the deployment is logged consistently. Oracle's pitch is that identity, networking and logging are already shared across the tenancy. For a compliance officer, that shared layer is the claim worth testing.
What to check before signing
Ask which EU region the inference runs in, and whether request data or model weights leave it. Ask whether the same model is available in every Oracle region or only some. Ask how traffic is routed when a provider changes a model version. Then ask for the price in EUR, with VAT, for your expected token volume, because the table above is not a quote.
We run local and hosted models on our own AI Arena hardware, and the pattern we keep running into is simple: swapping a model is easy, swapping a provider contract is not. Oracle is betting that enterprises would rather keep the contract and change the model.
Is Gemini in Oracle Cloud available to EU customers?
Oracle operates cloud regions inside the EU, and availability depends on the region you provision. Gemini is part of the catalogue Oracle sells worldwide, but regional availability and data-residency options have to be confirmed per service before you design around them.
Does buying Gemini through Oracle instead of Google Cloud change my AI Act duties?
The general-purpose model obligations rest with the provider that placed the model on the market. Your duties as a deployer depend on what you build, not on which reseller you bought the API from. The documentation trail does get longer when three or four vendors sit behind one application.
Is Oracle cheaper than buying directly from the model vendor?
Oracle does not publish a single catalogue rate and enterprise agreements are negotiated. The realistic comparison is your existing Oracle commitment against a new Google Cloud or OpenAI contract, including the procurement and legal hours each one consumes.
