Vertex AI pricing negotiation for Gemini in 2026. How to lower token, throughput and commit costs.

Vertex AI and Gemini spend is token and throughput pricing layered onto a Google Cloud commit. The deal turns on which meters you commit to and which you leave on demand.

How are Vertex AI and Gemini priced in 2026?

Gemini on Vertex AI bills per token for on demand inference, with input and output priced separately for each model. Provisioned throughput, training and platform tooling run on their own meters, all published on the Vertex AI pricing page. On demand tokens and provisioned throughput drive most of the spend we review.

Google now markets the Vertex AI platform as Gemini Enterprise Agent Platform. The rename changes nothing in the commercial mechanics, and invoices, quotes and many documentation pages still use the Vertex AI name, so we use it here too.

Output tokens cost 4 to 8 times as much as input tokens at current list prices, so workload shape matters as much as volume. A job that reads long documents and writes short answers costs far less per request than a drafting tool that writes pages from a short prompt.

Which pricing details change the bill most?

• Long context. On the Pro models, a prompt above 200,000 input tokens puts the whole request on a higher rate. Gemini 2.5 Pro goes to $2.50 input and $15 output per million tokens.
• Batch. Work submitted through the Batch API bills at 50 percent of the standard token price. Overnight extraction and classification jobs rarely need a real time answer.
• Priority tier. Google sells a priority tier for latency sensitive traffic at roughly 1.8 times the standard rate, for example $3.60 input and $21.60 output on Gemini 3.1 Pro Preview.
• Platform meters. Training, pipelines, vector search and the rest of the platform tooling each bill independently, and they rarely appear in the first AI proposal.

Treat published rates as the ceiling. What you actually pay depends on your commit, the discount instruments you choose and the alternatives you can put on the table.

What have we seen in recent Vertex AI and Gemini negotiations?

AI line items grew faster than any other Google Cloud meter and received the least scrutiny. That is the common thread across roughly 12 to 18 Google Cloud negotiations with material Vertex AI and Gemini components that I advised in 2024 and 2025.

Treat those ranges as benchmarks for what disciplined buyers achieved against the same Google sales approach. They are not promises. Your own meter history sets the baseline, and the first job in any Google AI negotiation is to produce it.

How does Vertex AI spend fit a Google Cloud commit?

Vertex AI consumption draws down your Google Cloud spend commitment, so the AI negotiation is a commit negotiation. That works in your favor. Growing AI usage can fill a platform commit you already hold and justify a better discount across all Google Cloud services.

Two instruments matter. The first is the enterprise commit in your Google Cloud agreement, where private pricing on specific SKUs is negotiated. The second is the set of committed use discount mechanics that apply to the rest of the platform, which Google has extended to generative AI.

What do Flexible Savings Plans do for Gemini spend?

Flexible Savings Plans are Google's spend based commitments for generative AI. They cover first party Gemini models, open source models sold as a service and participating third party models in Model Garden. A 1 year plan gives a 10 percent discount and a 3 year plan gives 20 percent, measured in dollars per month.

Read the terms before you sign one. You cannot cancel a plan, usage above the monthly amount bills at the on demand price, and provisioned throughput subscriptions and some fine tuning services draw down the plan without earning any extra discount. A 3 year plan also fixes your spend level while model prices keep falling.

How should you structure the AI component?

1. Measure current token and throughput burn by model and workload before any commitment conversation.
2. Commit platform wide at or below cleaned trailing spend and let AI growth fill the commit. See our guide to Google Cloud committed use discounts for how the tiers work.
3. Keep flexibility at the model level. Never commit to a single model family in a market that reprices every quarter.
4. Review provisioned throughput every quarter against measured load, never against launch forecasts.

How much can model routing cut Vertex AI token spend?

Routing is usually the largest single saving, and it needs no contract change. The price gap between Gemini tiers is wider than any discount Google will sign, so sending each workload to the cheapest model that passes your quality tests beats most rate negotiations.

Say a company sends 20 billion input and 4 billion output tokens a month, all to Gemini 2.5 Pro. At list price that is $25,000 for input and $40,000 for output, or $65,000 a month. Now route 40 percent of the traffic to Flash-Lite, 30 percent to Flash and keep 30 percent on Pro.

The routed bill is $25,740 a month, about 60 percent lower, and the annual difference is roughly $471,000. Moving the overnight Flash-Lite jobs to the Batch API would cut that line by another half.

When is provisioned throughput worth buying for Gemini?

Provisioned throughput is worth buying only for measured production load that needs stable latency and a predictable cost. Bought ahead of demand, it becomes shelfware, because you pay for the reserved capacity whether traffic arrives or not.

Google sells it in generative AI scale units (GSUs), each a fixed slice of model throughput, on four term lengths. From July 1, 2026, regional endpoints cost about 10 percent more than the global rates below for generally available Gemini 3 and later models.

The annual term looks cheapest, but idle capacity erases the gap. Twenty GSUs on a 1 year term cost $40,000 a month. If half sit idle for two quarters, $120,000 buys nothing. Ten GSUs on 1 month terms would have cost $27,000 a month over the same period, $13,000 less, with room to add units as load grows.

Why we disagree with locking in throughput early

The usual advice is to reserve provisioned throughput early because AI capacity is scarce and prices only rise. We disagree. In the 12 to 18 Google Cloud AI negotiations I advised in 2024 and 2025, per token prices for equivalent capability fell repeatedly as new Gemini tiers shipped. Early reservations sat half idle while better models arrived at lower rates.

Reserve only for measured production load, keep the right to shift reserved capacity to newer models, and let falling prices work for you. Scarcity arguments are a sales tactic, so ask Google to show the constraint in writing for your region. Our note on capacity scarcity in Google Cloud deals covers how to test the claim.

Commit to the burn you can measure. The market keeps lowering the price of every model you have not bought yet.

What lowers Google's AI pricing in a negotiation?

Three things lower Vertex AI and Gemini pricing: documented competitor quotes, measured burn data in place of growth forecasts, and routing discipline that proves you control consumption. Google's Cloud terms leave enterprise AI pricing open to negotiation inside your agreement. A quarterly throughput review clause adds a fourth tactic on the capacity side.

Why rival cloud quotes move Gemini rates

Google, Microsoft and AWS price frontier AI against each other week by week. A written quote from either rival, matched to your workload, shifts a Gemini rate faster than any volume argument. Get it for the same task, token mix and latency, using our guides to AWS Bedrock pricing and Azure OpenAI negotiation.

What will the Google account team say, and how should you answer?

• "Add your AI growth to the commit and we can move you up a discount tier." Answer that you will commit at trailing spend and let growth fill it, and ask what tier your measured spend earns today.
• "Capacity is tight, so reserve throughput now." Ask for the constraint in writing for your region and model, then reserve only your measured production peak on a short term.
• "A 3 year savings plan gets you 20 percent." Point out that the plan cannot be canceled while model prices fall every few months, and ask what the 1 year rate becomes with a model substitution right.
• "Our list price already beats OpenAI." Put the matched quote on the table and ask Google to price the same workload at the same token mix.

Which contract terms should a Vertex AI agreement include?

Ask for terms that keep your commitment flexible as models change. Each one below addresses a cost problem described earlier on this page. Put them in the enterprise agreement itself so they survive changes in the Google account team.

• Drawdown scope. A written list confirming that Gemini tokens, partner models in Model Garden, provisioned throughput and fine tuning all count toward the commit, so no AI line falls outside it.
• Model substitution. The right to apply commitments and reserved capacity to newer Gemini models at no penalty. Our swap and reallocation clause guide has sample wording.
• Price protection on successors. A new model tier that replaces one you use should cost no more per token than the tier it replaces, for the length of the term.
• Quarterly throughput review. The right to reduce or convert GSUs each quarter against measured load.
• Shortfall handling. Unused commit rolls into the next period, never billed as a penalty. The commit clause redlines guide covers the language.

How do you stop AI spend drifting after signature?

Set routing policy and quota governance at signature, before the first surprise invoice. AI spend drifts because every team can call a frontier model by default, and no contract term controls that.

• Default to small. Route routine classification and extraction to the cheapest Gemini tier that passes your quality checks.
• Quota by team. Give each project a budget and alerts on token meters from day one.
• Quarterly model review. Reprice workloads as new model tiers ship, because last year's routing table overpays today.

How do you check your own Vertex AI consumption?

Export Cloud Billing data to BigQuery and filter it by service and SKU. That gives you spend per model, per project and per meter, the twelve month history Google will ask for. Cloud Monitoring shows token counts and throughput usage per model, which you need to size any GSU purchase.

Cloud Billing budgets alert you but do not cap spending on their own. To stop a runaway project you need a programmatic response to the budget notification, or per project quotas. Our guide to Google Cloud cost allocation and tagging explains how to label AI workloads so each team sees its own bill.

What to do next

1. This month. Export twelve months of Vertex AI and Gemini meter data by model and project.
2. Before the first commit meeting. Build a routing table that maps each workload to the cheapest model tier that passes your quality tests.
3. Before any new order. Right size or cancel provisioned throughput that idles below measured production load.
4. Three months before renewal. Collect current written quotes from OpenAI and AWS Bedrock for matched workloads.
5. At the commit. Fold cleaned AI burn into the platform commit at or below trailing consumption, with drawdown scope and model substitution written in.
6. After signature. Set project level token quotas and a quarterly model repricing review. Our guide to negotiating Google Cloud AI contracts covers the wider agreement.

Frequently asked questions