Gemini 2.5 Flash: A Multimodal Cost Receipt

01 / The cheapest claim needs an actual receipt

A low input price is useful. A low cost per usable result is better. Gemini 2.5 Flash deserves consideration for budget multimodal workflows, but calling it the lowest-cost option everywhere would require a current, workload-specific comparison. Instead, let us price a concrete job: reading receipt images and returning short, structured answers.

The calculations below use Google's published standard API rates as a reference, not a quoted CheapRelay tariff or a promised discount. Google lists text, image and video input at $0.30 per million tokens, with output at $2.50 per million tokens. Confirm the applicable relay rates before funding production traffic. ([developers.googleblog.com](https://developers.googleblog.com/en/gemini-2-5-thinking-model-updates/?utm_source=openai))

02 / Receipt one: 10,000 images for $8.60

Assume each request contains 200 text tokens of instructions, an image accounting for 1,000 input tokens, and 200 billable output tokens. These are budgeting assumptions, not measured performance or a claim that every receipt image has the same token count.

• ■ Instructions: 200 × 10,000 = 2,000,000 tokens; reference cost $0.60.
• ■ Images: 1,000 × 10,000 = 10,000,000 tokens; reference cost $3.00.
• ■ Output: 200 × 10,000 = 2,000,000 tokens; reference cost $5.00.
• ■ Total: $8.60 for 10,000 requests, or $0.00086 per request.

This is calculated token spending, not an all-in operating budget. It excludes retries, paid tools, cache charges, storage and application infrastructure. At these reference rates, the output line costs more than the instructions and images combined. ([developers.googleblog.com](https://developers.googleblog.com/en/gemini-2-5-thinking-model-updates/?utm_source=openai))

03 / A longer answer turns $8.60 into $28.60

Keep the same inputs but increase billable output from 200 to 1,000 tokens per request. Output now totals 10 million tokens and costs $25.00. Add the unchanged $3.60 input bill, and the workload reaches $28.60: $20.00 more for the same 10,000 images.

For extraction, ask for merchant, date, currency, total and required line items rather than a narrative explaining the receipt. Test whether a compact response preserves the fields your application needs. The goal is not merely fewer tokens; it is fewer paid tokens per accepted record.

Include reasoning in that accounting. Google's output price includes thinking tokens. An additional 500 thinking tokens per request would add five million billable tokens across this workload, or $12.50 at the reference output rate. A short visible answer alone does not establish a small output bill. ([ai.google.dev](https://ai.google.dev/gemini-api/docs/pricing.md?utm_source=openai))

04 / Audio gets a separate line on the bill

Do not price every modality at the image-input rate. Google lists audio input for Gemini 2.5 Flash at $1.00 per million tokens. That difference belongs in your estimate before you compare workflows. ([ai.google.dev](https://ai.google.dev/gemini-api/docs/pricing.md?utm_source=openai))

Google documents audio tokenization at 32 tokens per second. Using that rule, a 60-second clip represents 1,920 audio tokens: $0.00192 in audio input. Add 200 billable output tokens costing $0.00050, and the illustrative total is $0.00242 per clip, or $2.42 for 1,000 clips, before text instructions and other charges. ([ai.google.dev](https://ai.google.dev/gemini-api/docs/tokens))

05 / Use CheapRelay without confusing reference and checkout prices

CheapRelay offers pay-as-you-go access through one API key, with no monthly fee or subscription. Its OpenAI-compatible routes use /v1, while Gemini-native routes use /v1beta. That gives an existing OpenAI SDK integration a practical connection path without making Google's reference prices a guarantee of the relay bill.

CheapRelay bills real token usage, including cache reads and writes. Keep those categories separate when reconciling spending; do not assume repeated instructions become free. For a pilot, record input, output, cache charges, retries and accepted results, then divide total spending by accepted results.

Sign up for CheapRelay and run a small receipt-extraction pilot before scaling. Choose Gemini 2.5 Flash as your lowest-cost multimodal option only when your measured cost per accepted result supports that label. The useful receipt ends with a verified total, not a superlative.