Google Gemini models: preview IDs, multimodal tokens and the 200K price boundary

Pro, Flash and Lite need different test material

For Pro, use a long specification with two conflicting requirements and ask for the conflict with section references. A broad summary can sound convincing while missing the exact contradiction you needed to find. For Flash Preview, use a scanned table and require a strict CSV, checking both character recognition and row alignment. For Flash-Lite, constrain a product description to a supplied fact sheet and check that attractive but unsupported claims never appear.

The 3.7 Flash entry is a separate identifier, not a silent replacement for the 3 Flash preview endpoint. Try it on a settings screenshot with visible labels and require navigation instructions that reference only controls present in the image. These are proposed exercises, not measured strengths. If two Gemini entries both pass your acceptance checks, compare completion time and billed usage over several representative inputs instead of assuming the model with the higher version number is the better purchase.

A 200K prompt changes Pro's price calculation

Google's standard Gemini 3.1 Pro Preview rate is $2 per million input tokens and $12 per million output tokens for prompts up to 200,000 tokens. Above that prompt boundary, the listed rates are $4 and $18. The threshold is about prompt length, not whether the answer happens to be long. An estimate that uses the smaller tier for a large uploaded dossier can therefore understate both sides of the bill.

For example, a 190,000-token prompt plus a 10,000-token answer costs $0.50 at the short-prompt standard rates. A 210,000-token prompt plus the same answer costs $1.02 at the longer-prompt rates. That is not a claim about real document sizes or measured output; it illustrates the discontinuity you should model when budgeting. Removing irrelevant appendices can matter more financially than shortening the final answer by a few sentences.

The standard text/image/video rates checked for Gemini 3 Flash Preview are $0.50 input and $3 output per million tokens; its audio input rate is $1. For Gemini 3.1 Flash-Lite, those figures are $0.25 input, $1.50 output and $0.50 audio input. We did not find an exact 3.7 Flash price in the reviewed rate card, so its row deliberately says unverified rather than borrowing a different Flash generation's price. Batch, priority and other service tiers are not used in these examples.

Gemini media tokens are not a word-count conversion

Google's token guide explicitly includes text, images and other non-text modalities in token accounting. The count_tokens operation counts input only; response usage can additionally distinguish output, thinking, cached content and tool-use tokens. A rough English characters-per-token estimate is useful for prose, but it cannot price a mixed-media request reliably. Count the actual request shape where supported and inspect the usage returned by the endpoint you use.

The documentation also separates media processing modes and model generations. Do not apply a static video tokens-per-second figure to an agentic processing flow, or copy a Gemini 4 context rule onto these Gemini 3 entries. System instructions contribute input tokens too. When comparing a photograph of a document with pasted OCR text, preserve the same evidence and check the real token result; fewer visible words on screen do not establish which representation costs less.

Provider multimodal support is not a promise that the chat interface exposes every transport or file type. Our table's vision flag comes from app configuration. Audio and video support described in Google's API documentation must still be checked against the workspace's actual upload and request path. If the task requires a live voice session, read the separate voice-pricing article rather than applying these text generation rates to minutes of conversation.

Treat preview availability as a deployment dependency

Two IDs here explicitly contain preview. Record the exact ID in an evaluation log and check availability in the intended Google project before relying on it for a recurring job. A working consumer Gemini session does not prove that the developer API project has the same entitlement, quotas or billing. Keep a small known-good input for an availability check, and avoid discovering an endpoint change halfway through an expensive document batch.

The configured million-token ceiling is useful for screening candidates, not a quality guarantee for long-context recall. Put a relevant fact near the beginning and another near the end of your trial document, then require both references. Keep an output allowance and check provider-generation rules rather than assuming that a maximum-sized input always leaves unlimited room for an answer. Review factual accuracy before comparing latency or the apparent fluency of the response.

Concrete Google model checks

Suggested evaluation protocols, not executed benchmark results.