Gemini 3.5 Flash API Pricing: Token Costs, Benchmarks, and Real Budget Examples
Compare Gemini 3.5 Flash API input, output, thinking, caching, and grounding costs with real workload examples, benchmark context, and practical ways to control spend.
Gemini 3.5 Flash costs $1.50 per 1 million input tokens and $9 per 1 million output tokens on Google’s standard paid tier. The output rate includes thinking tokens. Context-cache reads cost $0.15 per 1 million tokens, while stored cache content also carries a time-based storage charge.
On Poyo.ai, Gemini 3.5 Flash currently costs $0.90 per 1 million input tokens and $5.40 per 1 million output tokens. That is 40% below Google’s standard token rates, with pay-as-you-go credits and no monthly subscription requirement.
The headline prices are simple. The real bill is not. Agent loops, thinking tokens, repeated context, long answers, search grounding, and tool retries can make output and repeated input more important than the advertised input rate.
Pricing checked July 19, 2026. Google and Poyo.ai may update prices, features, or availability after publication.
Gemini 3.5 Flash API pricing at a glance
| Meter | Google standard paid tier | Poyo.ai |
|---|---|---|
| Input | $1.50 / 1M tokens | $0.90 / 1M tokens |
| Output, including thinking | $9.00 / 1M tokens | $5.40 / 1M tokens |
| Context-cache read | $0.15 / 1M tokens | Check current API terms |
| Context-cache storage | $1.00 / 1M tokens per hour | Check current API terms |
| Google Search grounding | 5,000 prompts shared across Gemini 3, then usage-based | Depends on endpoint and tool configuration |
Google also lists a free tier for Gemini 3.5 Flash. It is useful for evaluation, but production planning should use paid-tier prices, rate limits, data terms, and reliability requirements.
For current Poyo pricing and supported endpoints, see the Gemini 3.5 Flash API model page.
What counts toward a Gemini 3.5 Flash bill?
A useful first-pass formula is:
request cost =
input tokens × input rate
+ output and thinking tokens × output rate
+ cache storage
+ grounding or server-side tool chargesInput can include system instructions, conversation history, tool definitions, retrieved documents, images, audio, and video. Output includes the visible answer and billed thinking tokens. A request that looks short in the user interface may therefore contain a much larger metered payload.
Input tokens
Input is the least expensive standard token category, but it can become the largest line item when an application repeatedly sends long documents or complete conversation history. A 1 million-token context window is capacity, not a recommendation to fill every request.
Output and thinking tokens
Google’s $9 output rate explicitly includes thinking tokens. This matters because output is six times the standard input rate. A coding agent that explores several approaches, calls tools, and writes a long answer can spend more on reasoning and output than on the original prompt.
Context caching
Context caching can lower the read price for repeated material such as a stable system prompt, product catalog, policy manual, or codebase summary. It is most useful when the same cached block is reused enough times to offset storage charges and cache-management complexity.
Search grounding and tools
Google’s pricing page lists an included monthly allowance for Search grounding, followed by a per-query charge. One model request can produce more than one search query. Tool-heavy applications should log tool calls separately from token usage so the source of a cost increase is visible.
How much does Gemini 3.5 Flash cost in practice?
The examples below use standard token pricing and exclude taxes, special processing tiers, and separate tool charges.
Example 1: a small coding assistant
Suppose a coding assistant consumes 10 million input tokens and 2 million output tokens per month.
| Provider | Input cost | Output cost | Monthly total |
|---|---|---|---|
| Google standard | $15.00 | $18.00 | $33.00 |
| Poyo.ai | $9.00 | $10.80 | $19.80 |
The output is only one-fifth of the input volume, yet it costs more than the input on both routes. This is why output limits and reasoning settings deserve attention.
Example 2: a support agent with repeated context
Imagine 100,000 monthly conversations. Each starts with 8,000 tokens of policies and product context, then produces 1,000 output tokens.
Without caching or retrieval:
- Input: 800 million tokens
- Output: 100 million tokens
- Google standard: $1,200 input + $900 output = $2,100
- Poyo.ai token rates: $720 input + $540 output = $1,260
This design repeatedly pays to send the same context. A better architecture would cache the stable prefix, retrieve only relevant policy sections, or prepare a smaller structured summary.
Example 3: a multimodal document workflow
For PDFs, images, audio, and video, do not estimate cost from file size alone. Measure the input tokens returned by the API for representative files. Then test short, medium, and difficult cases because media length and model reasoning can change both sides of the bill.
A practical pilot should record:
- input tokens by media type;
- output and thinking tokens;
- tool and grounding calls;
- latency;
- task success without human correction;
- retries per successful task.
Why a Gemini 3.5 Flash bill can exceed the estimate
Thinking is billed as output
If an application enables deeper reasoning for every request, it pays the output rate even when the visible response is short. Use the lightest thinking level that still meets the task’s quality bar.
Agent loops multiply both input and output
Each tool result may return to the model as new input. Failed calls, ambiguous tool schemas, and missing stop conditions can create expensive loops. Set iteration limits and capture the reason for every retry.
Full context is resent unnecessarily
Sending an entire repository, conversation, or document collection on every turn is rarely the most economical design. Summaries, retrieval, cacheable prefixes, and state stored outside the model can reduce repeated input.
Unbounded answers are expensive
Output is the premium token category. Set realistic maximum output values, ask for structured concise responses, and generate long reports only when users actually need them.
Grounding can fan out
A single request may trigger multiple search queries. Applications that require predictable unit economics should set tool policies and monitor query counts rather than treating grounding as free context.
Gemini 3.5 Flash benchmarks in context
Google positions Gemini 3.5 Flash as an agentic and coding model that preserves Flash-class speed. Its May 2026 launch data reports:
| Evaluation | Gemini 3.5 Flash result | What it tests |
|---|---|---|
| Terminal-Bench 2.1 | 76.2% | Command-line and agentic coding execution |
| GDPval-AA | 1,656 Elo | Professional knowledge-work performance |
| MCP Atlas | 83.6% | Agent and tool-use capability |
| CharXiv Reasoning | 84.2% | Multimodal chart and visual reasoning |
Google also reports that Gemini 3.5 Flash produces output at roughly four times the speed of other frontier models in its comparison. These figures are useful signals, not a substitute for application testing. Harness design, tool definitions, reasoning settings, latency targets, and prompt quality all affect real performance.
The important economic question is not “which benchmark is highest?” It is:
How much does one successful task cost after tokens, tools, retries, and human corrections?
Gemini 3.5 Flash compared with alternative models
| Model | Best fit | Main cost question |
|---|---|---|
| Gemini 3.5 Flash | Fast multimodal agents, coding, tool workflows | Are thinking and repeated context controlled? |
| GPT-5.6 Terra | Balanced coding and production reasoning | Does stronger task completion offset a higher per-token tier? |
| Claude Sonnet 5 | Agentic coding, computer use, long professional workflows | How does the newer tokenizer affect equivalent-request cost? |
| Gemini 3.1 Flash-Lite | High-volume simpler processing | Does the task really require frontier-level reasoning? |
Use this as a routing shortlist. Run the same representative workload through each candidate before choosing a default.
Seven ways to reduce Gemini 3.5 Flash API costs
- Set an output ceiling. Avoid paying for explanations or repeated context that the product does not show.
- Adjust thinking by task difficulty. Reserve deeper reasoning for requests that benefit from it.
- Cache stable prefixes. Reuse system prompts or reference material when the hit rate justifies storage.
- Retrieve before generating. Send the relevant document sections instead of the entire corpus.
- Limit agent iterations. Stop or escalate after a defined number of tool failures.
- Route simple tasks down. Classification and extraction may fit a lighter model.
- Track cost per successful task. Token cost alone hides retries and correction work.
How to call Gemini 3.5 Flash on Poyo.ai
Poyo.ai supports the model ID gemini-3.5-flash. Teams can use an OpenAI-compatible chat endpoint or the Gemini-native format documented for the model.
curl https://api.poyo.ai/v1/chat/completions \
-H "Authorization: Bearer $POYO_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "gemini-3.5-flash",
"messages": [
{
"role": "user",
"content": "Review this agent plan and identify the three most expensive failure modes."
}
]
}'Create an API key from the Poyo dashboard, then check the Gemini 3.5 Flash model page and linked API documentation before production deployment.
Is Gemini 3.5 Flash worth the price?
Gemini 3.5 Flash is a strong candidate when an application needs fast coding, tool use, multimodal input, or long-context reasoning. It is less compelling when a request is predictable enough for a smaller model.
The best production pattern is usually selective routing:
- use a lighter model for simple, repetitive tasks;
- use Gemini 3.5 Flash for requests that need stronger reasoning or multimodal context;
- escalate only the hardest cases to a more expensive frontier tier.
That approach captures the model’s speed and capability without paying frontier-model costs for every request.
Frequently asked questions
How much does Gemini 3.5 Flash cost per million tokens?
Google’s standard paid tier is $1.50 per 1 million input tokens and $9 per 1 million output tokens, including thinking tokens. Poyo.ai currently lists $0.90 input and $5.40 output per 1 million tokens.
Is the Gemini 3.5 Flash API free?
Google lists a free tier, but production applications should evaluate its rate limits and data terms. Poyo.ai uses pay-as-you-go credits.
Are Gemini 3.5 Flash thinking tokens charged?
Yes. Google’s pricing table states that the output rate includes thinking tokens.
Does Gemini 3.5 Flash support context caching?
Yes. Google lists paid cache-read and storage rates. Caching is most economical for stable context reused across enough requests.
What is the Gemini 3.5 Flash model ID?
The model ID is gemini-3.5-flash.
Can Gemini 3.5 Flash use an OpenAI-compatible API?
Yes. Poyo.ai exposes an OpenAI-compatible chat-completions route as well as a Gemini-native integration path.
