Gemini 3.8 Flash API cost
At list prices from Google (Gemini API), Gemini 3.8 Flash costs $0.75 per million input tokens and $3.75 per million output tokens, with cached input at $0.075. Across the standard workloads below it would cost from $188.20 a month (long-form writing) to $2,099 a month (cached-prompt assistant).
It is the 6th cheapest of 13 models at 4 workloads and 5th cheapest of 12 for the cached-prompt assistant workload. Output tokens are 91% of the bill for long-form writing but 7.7% for long-document questions. Prompt caching takes the bill for the cached-prompt assistant workload from $4,563 to $2,099 a month, 54% less. Sent through the batch API, the short chat workload would cost $285.16 a month instead of $570.31.
Prices on this page last verified on Sources
Published prices
Google (Gemini API) list prices per million tokens. In our mid-range tier.
| Prompt length | Input | Cached input | Output | Batch input | Batch cached input | Batch output |
|---|---|---|---|---|---|---|
| All prompt lengths | $0.75 | $0.075 | $3.75 | $0.375 | $0.0375 | $1.875 |
Cost at five standard workloads
The same traffic on every model, so the rank shows where Gemini 3.8 Flash sits. “vs cheapest” compares it with the cheapest model for that workload.
| Workload | Requests a day | Tokens in / out | Per 1,000 requests | Per month | Via batch API | Rank, all models | Rank, mid-range tier | vs cheapest |
|---|---|---|---|---|---|---|---|---|
| Short chat | 10,000 | 1,000 / 300 | $1.88 | $570.31 | $285.16 | 6th of 13 | 1st of 3 | 7.5× (Claude Haiku 5.5) |
| Long-document questions | 1,000 | 30,100 / 500 | $24.45 | $743.69 | $371.84 | 6th of 13 | 1st of 3 | 7.5× (Claude Haiku 5.5) |
| Bulk tagging | 200,000 | 400 / 10 | $0.337 | $2,053 | $1,027 | 6th of 13 | 1st of 3 | 7.5× (Claude Haiku 5.5) |
| Long-form writing | 500 | 1,500 / 3,000 | $12.37 | $188.20 | $94.10 | 6th of 13 | 1st of 3 | 7.5× (Claude Haiku 5.5) |
| Cached-prompt assistant | 20,000 | 8,000 / 400 | $3.45 | $2,099 | $1,049 | 5th of 12 | 1st of 3 | 7.5× (Claude Haiku 5.5) |
Equal costs share a rank. “Via batch API” prices the same traffic with every request sent through the provider’s batch API, using its published batch prices.
Workload assumptions
Short chat
Short questions and answers at moderate volume.
| Part of the prompt | Tokens | Cached | Why |
|---|---|---|---|
| Instructions | 300 | No | Short instructions, treated as uncached. |
| Message and recent turns | 700 | No | Changes with every request. |
| Total | 1,000 | None cached |
- Requests a day: 10,000
- A popular in-app assistant.
- Output tokens per request: 300
- A reply of about 220 words.
- Batch API: not used
- Real-time replies.
These are stated assumptions for illustration, not measurements. Cached parts are assumed to be read from the prompt cache on every request. Token counts are the same for every model, although providers count tokens differently. Use the calculator to try your own figures.
Long-document questions
Long input with a short answer: most of the cost is input tokens.
| Part of the prompt | Tokens | Cached | Why |
|---|---|---|---|
| Document | 30,000 | No | About 22,000 words, different every time. |
| Question | 100 | No | One or two sentences. |
| Total | 30,100 | None cached |
- Requests a day: 1,000
- A research or legal team asking questions of whole documents.
- Output tokens per request: 500
- An answer of about 370 words.
- Batch API: not used
- Real-time answers.
These are stated assumptions for illustration, not measurements. Cached parts are assumed to be read from the prompt cache on every request. Token counts are the same for every model, although providers count tokens differently. Use the calculator to try your own figures.
Bulk tagging
Very high volume with a tiny output: a label per item.
| Part of the prompt | Tokens | Cached | Why |
|---|---|---|---|
| Instructions and labels | 300 | No | Short instructions, treated as uncached. |
| Item | 100 | No | A short text. |
| Total | 400 | None cached |
- Requests a day: 200,000
- Items such as tickets, reviews or records tagged as they arrive.
- Output tokens per request: 10
- One label.
- Batch API: not used
- Tagged as items arrive.
These are stated assumptions for illustration, not measurements. Cached parts are assumed to be read from the prompt cache on every request. Token counts are the same for every model, although providers count tokens differently. Use the calculator to try your own figures.
Long-form writing
A short brief and a long answer: most of the cost is output tokens.
| Part of the prompt | Tokens | Cached | Why |
|---|---|---|---|
| Brief and style notes | 1,500 | No | Different for each piece. |
| Total | 1,500 | None cached |
- Requests a day: 500
- Articles, reports or long emails drafted a day.
- Output tokens per request: 3,000
- About 2,200 words.
- Batch API: not used
- Drafts are wanted straight away.
These are stated assumptions for illustration, not measurements. Cached parts are assumed to be read from the prompt cache on every request. Token counts are the same for every model, although providers count tokens differently. Use the calculator to try your own figures.
Cached-prompt assistant
A large fixed prompt read from the prompt cache on every request.
| Part of the prompt | Tokens | Cached | Why |
|---|---|---|---|
| Instructions and tool definitions | 6,000 | Yes | The same on every request, so read from the prompt cache. |
| Conversation and request | 2,000 | No | Changes with every request. |
| Total | 8,000 | 75% cached |
- Requests a day: 20,000
- A busy assistant with long instructions and tool definitions.
- Output tokens per request: 400
- A reply of about 300 words.
- Batch API: not used
- Real-time replies.
These are stated assumptions for illustration, not measurements. Cached parts are assumed to be read from the prompt cache on every request. Token counts are the same for every model, although providers count tokens differently. Use the calculator to try your own figures.
Your traffic looks different? Enter your own figures in the calculator.
Related pages
Head-to-head comparisons
Other models in the same tier
Sources
Sources: aistudio.google.com (checked )
- Tier: Provider text: "Our most intelligent Flash model, engineered for long-horizon software engineering, autonomous agents, and complex enterprise workflows." Most capable model of the faster Flash line: mid. aistudio.google.com (checked )
Prices are list prices as each provider publishes them, used unchanged. How the figures are calculated: methodology.