Google Gemini API Pricing 2026: Every Model and Cost Explained

Google's Gemini API has become one of the most cost-competitive ways to build with a frontier model, but the pricing page is easy to misread. Rates change with each model, output costs several times more than input, and the newer reasoning models charge for the tokens they spend thinking. A small change in which model you call can swing your bill by 10x.

That makes Gemini pricing less of a sticker question and more of a FinOps one. For anyone building a product on the API, the real cost depends on model choice, prompt size, output length, and whether you use features like context caching and Search grounding.

This guide breaks down Gemini API pricing by model for 2026, explains the tiered pricing on Pro, shows what actually drives your bill, compares Gemini against OpenAI and Anthropic, and lays out practical ways to cut costs without losing quality.

Key Takeaways

The Gemini API is billed per million tokens on a standard paid tier, with a free tier (Flash and Flash-Lite) for prototyping; Pro models have been paid-only since April 1, 2026.

Gemini 3.1 Pro is $2 input and $12 output per million up to 200K tokens, and $4 and $18 above that; it supports a 2-million-token context window, the largest in production.

The current Gemini 3.x Flash models are about $0.75 and $3.75 on introductory pricing through December 31, 2026, then $1.50 and $7.50; Flash-Lite tiers run $0.10 to $0.30 input.

Every Gemini model includes a 1-million-token context window, even the cheapest, which is unusual among frontier providers.

Thinking tokens are billed at output rates, batch is 50 percent off, and cached input reads cost about 10 percent of the base rate, so real costs depend heavily on how you use it.

How Gemini API pricing works

Gemini is billed per token, not per request. A token is a chunk of text, roughly four characters or three-quarters of a word in English. Every call has two metered parts:

• Input tokens. Everything you send: the prompt, system instructions, conversation history, and any files or images, counted as tokens.
• Output tokens. Everything the model returns. On the Gemini 2.5 reasoning models, this also includes thinking tokens, the internal reasoning the model generates before its final answer.

Output is the expensive side. On most models, output tokens cost several times more than input, so a chatty model that reasons at length can cost far more than the input rate suggests. There are also two billing surfaces for the same models: the Gemini API through Google AI Studio, and Vertex AI for enterprise governance. Per-token rates are broadly similar; Vertex adds enterprise controls and its own commitments.

Free tier: Google AI Studio offers a free tier with lower rate limits, useful for prototyping. On the free tier, your prompts and responses may be used to improve Google's products. On the paid tier, your data is not used for training. If you handle anything sensitive, build on the paid tier.

Gemini consumer plans

Gemini API pricing by model (2026)

Here is the per-model breakdown for the Gemini 2.5 family, the current generation. Rates are per million tokens for standard text prompts. Confirm live numbers on Google's Gemini API pricing page, since Google updates models and rates regularly.

The pattern to internalize: Pro is for the work that genuinely needs it, Flash handles most production traffic, and Flash-Lite is for cheap, high-volume jobs where a smaller model is good enough. Choosing the right model per task, rather than defaulting everything to Pro, is the single biggest lever on your bill.

Gemini 3.1 Pro pricing explained

Gemini 3.1 Pro uses tiered pricing based on how large your prompt is. Short and medium prompts bill at the lower rate; once a single prompt crosses roughly 200K tokens, both input and output step up to a higher rate. It also supports a 2-million-token context window, the largest of the major flagships.

Two things surprise teams here. First, the jump is per prompt, so a few very large requests can quietly bill at double the rate while your dashboards still show an average. Second, the output price includes thinking tokens. A hard reasoning task can generate far more internal thinking than visible answer, so a 500-token reply might bill as several thousand output tokens. If you do not need deep reasoning on a given call, a Flash model or a lower thinking budget usually wins on cost.

Gemini 3.x Flash and Flash-Lite: the cost-efficient tiers

Most production traffic does not need the Pro model. The Flash tiers exist precisely so you do not overpay for routine work. The current Gemini 3.x Flash models are the balanced choice, priced at about $0.75 input and $3.75 output per million on introductory pricing through December 31, 2026 (then $1.50 and $7.50), with strong quality well below Pro on output. They are the right default for chat, summarization, extraction, and most app features, and because they are thinking models the same thinking-token caution applies, so tune the thinking budget down on simple calls.

Gemini 3.5 Flash-Lite ($0.30 input, $2.50 output) and 3.1 Flash-Lite ($0.25 and $1.50) are the budget tiers, built for high-throughput, latency-sensitive work like classification, routing, tagging, and simple transformations. The legacy Gemini 2.5 Flash-Lite at $0.10 and $0.40 remains the cheapest option only until it retires on October 16, 2026, so new work should target the 3.x tiers.

What actually drives your Gemini bill

Per-token rates are only the starting point. These five factors decide what you actually pay:

• Output volume and thinking tokens. Output costs several times more than input, and on 2.5 models thinking tokens count as output. Long, verbose, or heavily reasoned responses are where cost concentrates.
• Prompt size. Large context windows are powerful but you pay for every token you stuff in. Resending long histories or whole documents on each call adds up fast, and on Pro can tip you into the higher pricing tier.
• Context caching. If you reuse the same large context across many calls, caching lets you pay a reduced rate for the cached tokens plus a storage fee per hour. Used well it saves money; left running it becomes a quiet recurring charge.
• Search grounding. Grounding responses with Google Search has its own cost: a free daily allowance, then a per-thousand-requests charge. Heavy grounded traffic is a separate line item from token usage.
• Multimodal inputs. Images, audio, and video are converted to tokens and billed accordingly, and audio input is priced higher than text on most models. Video and audio workloads cost more than they look.

Worked example: estimating a monthly Gemini bill

Token math feels abstract until you put numbers to it. Here is a simple estimate for a product that handles 500,000 requests per month, where each request sends about 2,000 input tokens and receives about 500 output tokens, including thinking. The same workload costs very differently depending on the model you route it to.

The same traffic ranges from roughly $200 to $3,750 a month depending purely on model choice. That 18x spread is the entire argument for routing carefully and measuring cost per feature, not just total spend. It also shows why output dominates: on Flash and Pro, the output line is larger than the input line even though there are far fewer output tokens.

How to cut your Gemini API costs

Most Gemini bills can be reduced meaningfully without hurting quality. In rough order of impact:

1. Route by task, not by habit. Default to Flash or Flash-Lite and reserve Pro for the requests that genuinely need it. Model routing is the largest single lever in almost every deployment.
2. Control output length and thinking. Cap output tokens, ask for concise answers, and lower the thinking budget on simple calls. Since output dominates cost, this pays off immediately.
3. Use Batch mode for offline work. Non-urgent jobs run through the Batch API typically cost about half the interactive rate. For anything that does not need a real-time response, this is close to free money.
4. Cache reused context. If many calls share the same long instructions or documents, context caching cuts the per-call cost of those repeated tokens. Just remember the storage fee and clear caches you no longer need.
5. Trim prompts and histories. Send only the context the model needs. Summarize long conversation histories instead of resending them in full, and avoid pasting entire documents when a relevant excerpt will do.

Gemini vs OpenAI vs Claude API pricing

Gemini's headline advantage is price, especially on the Flash tiers. Here is how the flagship and workhorse models compare on published list rates. Treat these as approximate, since every provider updates pricing often.

On raw token price, Gemini Flash sits well below the OpenAI and Anthropic flagships, which is why high-volume products often standardize on it. The flagship comparison is closer once you account for quality differences per task. The practical takeaway is the same one that applies to every provider: the cheapest model that meets your quality bar wins, and what matters most is total AI spend visibility across all the models you run, not any single sticker price.

5 hidden costs and gotchas to watch

The pricing table rarely tells the whole story. These are the surprises that show up on the invoice:

• Thinking tokens inflate output. On 2.5 reasoning models, internal thinking bills as output. A short visible answer can carry a large hidden output cost.
• The Pro tier jump is per prompt. A handful of very large prompts can bill at the higher Pro rate while averages look fine.
• Cache storage keeps charging. Context caching saves on repeated tokens but adds an hourly storage fee. Forgotten caches become a slow leak.
• Search grounding is separate. Grounded requests are billed apart from tokens once you pass the free daily allowance.
• No native cost allocation. Billing does not split spend by team, feature, model, or customer. At any real scale, you cannot answer cost per customer or cost per feature without extra tooling.

Conclusion

Gemini is one of the most affordable ways to build with a frontier model, but the bill is driven by model choice, output length, and features like caching and grounding, not the headline rate.

Route work to the cheapest model that clears your quality bar, watch output and thinking tokens, and keep AI spend allocated. In 2026, visibility is the real cost lever.