PPROMPT WORKBOOK Build a prompt
Model guide · checked 23 September 2026

Prompting Gemini

Gemini 3.8 Flash is Google's current stable all-round model for code, agents and multimodal input. Flash‑Lite is the volume option, while Pro remains a preview model.

Choose by release status, not just model name

ModelAPI input / outputChoose it when
Gemini 3.8 FlashStable, GA$0.75 / $3.75through 31 December 2026you need a strong multimodal production model for code, agents and long inputs
Gemini 3.1 Pro PreviewPreview$1 / $5through 31 December 2026you are experimenting with advanced multimodal understanding and can accept shorter deprecation notice
Gemini 3.5 Flash‑LiteStable$0.30 / $2.50you run translation, simple data processing or many small agent tasks

Gemini 3.8 Flash has a 1.05M token context window, up to 64,000 output tokens and low, medium and high thinking levels. Its introductory price rises to $1.50 / $7.50 on 1 January 2027. New projects should normally begin with 3.8 and keep an older generation only where their own regression tests require it.

Specify input, constraints and response shape

Separate the task from the material it concerns. State the kind of answer you need, and use the API's structured output capability when another system must read the result.

Use consistent examples

Few-shot examples can steer format, wording and pattern. Keep formatting consistent and vary the content enough that the model learns the rule rather than accidental details.

Put long material first and the question last

Google recommends placing large documents, code, video or datasets before the specific question and referring explicitly to the relevant modality. Critical rules may live in the system instruction or early in the prompt.

Use thinking level, not chain-of-thought formulas

Choose a lower thinking level for simple extraction or classification and increase it for complex agent runs when evaluations show a gain. Thinking tokens may be billed as output. Split the process into multiple API stages when an intermediate result must be inspected, logged or approved.

Separate format from action

Structured Outputs control final format. Function Calling describes an action or intermediate operation. Valid JSON does not guarantee true values, and tool arguments must be validated before execution.

Parameter change: newer Gemini generations move more control into the model and thinking level. Do not copy old recipes for temperature, top‑p, top‑k or a prefilled model turn without testing them.

Sources