Gemini 3.8 Flash vs 3.1 Pro: Cost per Answer

2026-10-25 Model price-offs EN

01 / The rate card behind each answer

A 2,000-input-token request with 500 billable output tokens costs $0.003375 on Gemini 3.8 Flash at Google's introductory rates. That is a calculated example, not a measured average. Google lists $0.75 per million input tokens and $3.75 per million output tokens through December 31, 2026. ([ai.google.dev](https://ai.google.dev/gemini-api/docs/whats-new-gemini-3.8?utm_source=openai))

For Gemini 3.1 Pro, Google's published rates for prompts below 200,000 tokens are $2 per million input tokens and $12 per million output tokens. The same example therefore costs $0.01. These are direct-provider reference prices, not verified CheapRelay quotes. ([ai.google.dev](https://ai.google.dev/gemini-api/docs/gemini-3?authuser=19&hl=en))

Every receipt below assumes standard text requests, no cache hits, no external tools, and fixed billable token counts. Taxes and other charges are excluded. The purpose is to make the arithmetic inspectable, not to imply identical answer quality.

02 / Receipt one: a short support answer

Imagine an application sending instructions, conversation history, and a customer question totaling 2,000 input tokens. Budget 500 billable output tokens for the response.

■ Gemini 3.8 Flash input: 2,000 / 1,000,000 × $0.75 = $0.001500.
■ Flash output: 500 / 1,000,000 × $3.75 = $0.001875.
■ Flash total: $0.003375 per answer; $3.375 per 1,000 answers.
■ Gemini 3.1 Pro input: 2,000 / 1,000,000 × $2 = $0.004000.
■ Pro output: 500 / 1,000,000 × $12 = $0.006000.
■ Pro total: $0.010000 per answer; $10 per 1,000 answers.

Calculated from those published rates, Flash costs 66.25% less for this exact token mix. That percentage describes the receipt, not a universal discount or performance advantage. ([ai.google.dev](https://ai.google.dev/gemini-api/docs/whats-new-gemini-3.8?utm_source=openai))

03 / Receipt two: answering over a larger document

Now budget 20,000 input tokens and 2,000 billable output tokens. This hypothetical request could represent a document question with a longer explanation.

■ Flash: $0.015 input + $0.0075 output = $0.0225.
■ Pro: $0.040 input + $0.024 output = $0.0640.
■ For 10,000 such requests: $225 versus $640.

The calculated difference is $415 across that workload. Notice that the savings percentage changes with the input/output mix; there is no single price per answer independent of token consumption. ([ai.google.dev](https://ai.google.dev/gemini-api/docs/whats-new-gemini-3.8?utm_source=openai))

04 / The receipt must include thinking and failed attempts

Google's output pricing includes thinking tokens. A 500-token visible reply is therefore not necessarily a 500-token billable output. Use billing usage rather than estimating cost from the text displayed on screen. ([ai.google.dev](https://ai.google.dev/gemini-api/docs/pricing))

For your own comparison, divide total test spending by accepted answers. Define acceptance before testing: correct fields, valid formatting, supported claims, or passing application checks. Keep unsuccessful attempts in the numerator. Otherwise, a cheap request can look like a cheap result when it is not.

For cached workflows, request a separate accounting of fresh input, cache reads, cache writes, output, and any additional charges. Do not apply the uncached receipts mechanically to a different billing arrangement.

05 / The January 2027 price change alters the budget

Google lists Flash standard pricing of $1.50 per million input tokens and $7.50 per million output tokens starting January 1, 2027. Receipt one would become $0.00675; receipt two would become $0.045. Those are scheduled-rate calculations, not current introductory prices. ([ai.google.dev](https://ai.google.dev/gemini-api/docs/pricing))

06 / Use CheapRelay billing to build your actual receipt

CheapRelay offers pay-as-you-go access to both models through one API key, with billing per real token, including cache reads and writes, and no monthly fee or subscription. Check its current model-specific rates before substituting them into these examples; this article does not establish a relay discount.

Sign up for CheapRelay, run a small representative test, and compare total billed spending per accepted answer. Keep the cheaper model where it meets your acceptance criteria, and pay more only where your own results justify the difference.

More reading

Date	Aisle	Written in	Title
2026-10-25	Model price-offs	EN	Small Routers, Big Answers: A $652.80 AI Receipt A transparent routing budget: spend $4.80 screening 100,000 requests, avoid unnecessary answer calls, and check the break-even math.
2026-10-25	Model price-offs	EN	The Receipt Behind a 12-Call Tool Agent A concrete cost breakdown for tool-heavy agents, showing how context growth, model routing, tool fees, and retries change the bill.
2026-10-25	Model price-offs	EN	Gemini 4 Argon vs Claude Opus 5: The Receipt A token-by-token cost comparison for reasoning-heavy prompts, including thinking, caching, retries, and the price of an accepted answer.
2026-10-25	Model price-offs	EN	The Model Ladder: A $10.24 AI Receipt A worked model-routing budget: start cheap, validate results, and escalate failures instead of paying premium rates for every request.
2026-10-25	Model price-offs	EN	Hunyuan 4 vs MiMo V2.6 Pro: A Light-Use Receipt Compare light-workload token receipts, cache scenarios and retry costs using published rates, without confusing provider prices with relay billing.
2026-10-25	Model price-offs	EN	Compare Models by Cost per Accepted Answer A receipt-style example showing how tokens, caching, retries, and acceptance rates change the real cost of an AI answer.