A Gemini Batch job can carry a lower token rate yet still have separate cached-context storage and search-grounding costs. Model version and prompt length must be chosen before calculating the discount.
Pricing snapshot: Vendor first-party documentation checked on 25 October 2026. Original price effective date unknown unless separately declared. Amounts are published list figures, subject to tier, region, currency and eligibility.
Exact published context bands
Google’s Gemini Developer API pricing schedule lists Gemini 3.1 Pro Preview with different token tariffs depending on whether prompts are at most 200,000 tokens or above that threshold. Its Standard rate for short prompts is $2 per million input tokens and $12 per million output tokens including thinking; the corresponding Batch rows are $1 and $6. Above 200,000 tokens, the published Batch input rate is $2 and output is $9 per million. These are whole-request context categories, not merely a few expensive excess tokens.
Google identifies this as a preview model. A preview rate and rollout cannot be generalized to a future stable version without additional first-party evidence. Use the exact model ID and tier in pricing records.
The batch queue is not an interactive service
The official Batch API guide describes asynchronous jobs and says eligible usage is commonly discounted relative to standard interactive usage. A job may complete faster than its advertised service objective, but a planning system must allow for longer turnaround and monitor failures. Do not schedule urgent customer replies through Batch just to halve a token line item.
Store submission and completion status, retries, token consumption and each batch’s declared target model. Rate-limiting and job-size constraints may differ from interactive calls; a monthly ledger that assumes identical throughput is not a capacity plan.
A threshold-aware worked example
Consider an illustrative Batch request with 100,000 billable uncached input tokens and 10,000 output tokens, no tools, no cache storage. At the cited short-context Batch rates of $1 and $6 per million, model cost is $0.10 plus $0.06 = $0.16. For a distinct 250,000-input-token request and 10,000 output tokens, select the long-context Batch rates before multiplying: $0.50 + $0.09 = $0.59.
This does not establish how much real research or media processing costs, because input modalities and separately billed storage or grounding activity may be present.
Context caching has a storage meter
Google’s tariff separates cached-input usage from storage time, expressed in token-hours. This means keeping a large cache alive may have a continuing cost even while no new prompt is processed. A simplistic cash model that applies only a discounted cache-hit rate will understate long-lived caches. Track cache size, start/end time, read volume, lifespan and per-model storage rate explicitly.
A batch request reusing cached content also needs to follow the documented compatibility rules. The guide states that cached content can be referenced in certain Batch requests, but a hit does not imply the cache-store interval is free.
Thinking, tool calls and eligibility
Output-token pricing includes thinking tokens on Gemini 3.1 Pro. A short final answer does not prove that a reasoning-heavy request used only a handful of output tokens. Google Search grounding has its own charge model and quota; it should be added only when enabled and executed. Region, free-tier availability and service mode can change eligibility, and a public global US-dollar tariff is not necessarily a local tax-inclusive price.
Do not add a blanket 50% discount to a rate already labeled Batch. For developer documentation, link the record to the checked pricing page and avoid extrapolating to Vertex AI or Google Workspace subscriptions.
Pricing evidence boundary
This cost discussion concerns Google (gemini-3.1-pro-preview). Reproduce a quote only with its exact model or plan identifier, chargeable unit, billing region, processing mode and timestamp; a changing vendor rate card is not a permanent price guarantee. If any one of these dimensions is missing, keep the associated rate and worked scenario under editorial review instead of filling the gap with an estimate.
Sources & Further Reading
- ai.google.dev — pricing (official documentation; accessed 2026-10-25)
- ai.google.dev — batch api (official documentation; accessed 2026-10-25)