Gemini API cost: two prices double before January

If you picked Gemini 4 Argon this year because it cost the same as Claude Sonnet, check Google's pricing page again. On December 1, Argon goes from $2 to $4 per million input tokens and from $10 to $20 per million output tokens. Same model, same traffic, twice the invoice.

Gemini 3.8 Flash follows on January 1: $0.75 to $1.50 input, $3.75 to $7.50 output.

Price increases on an API are rare enough that most cost estimates don't plan for them. So let's take one concrete workload, price it on every current Gemini model before and after the change, and then look at the four things that move a Gemini bill more than the price list does: thinking tokens, caching, the Batch API and grounding.

I maintain a Gemini price table for a cost calculator, so I re-check Google's pricing page every few weeks. In October, two rows had something I hadn't seen on an LLM price list before: a second price with a start date. No announcement in my inbox, no deprecation notice, just a footnote on the page.

• A document pipeline of 300,000 requests a month costs $6,000 on Gemini 4 Argon today and $12,000 from December 1. On Gemini 3.8 Flash: $2,250 today, $4,500 from January 1.
• Thinking tokens bill as output. 1,500 thinking tokens per request add 75% to the Flash bill, and they don't appear in the answer.
• The Batch API cancels the increase exactly (50% off) for work that can wait. A cached 5,000-token prefix cuts another 45–48%. Gemini 3.1 Flash-Lite runs the same job for $780.

How is the Gemini API billed?

Per million tokens, input and output separately. Cached input costs 10% of the input price, the Batch API halves input and output, and Pro models charge more once a prompt passes 200,000 tokens. These are the paid-tier list prices from Google's pricing page, in USD per million tokens, checked October 14, 2026:

Model	Input	Output	Cached input	Scheduled change
Gemini 4 Argon	$2	$10	$0.10	$4 / $20 from Dec 1, 2026
Gemini 3.1 Pro Preview	$2	$12	$0.20	$4 / $18 over 200K tokens
Gemini 3.8 Flash	$0.75	$3.75	$0.075	$1.50 / $7.50 from Jan 1, 2027
Gemini 3.5 Flash-Lite	$0.30	$2.50	$0.03	none
Gemini 3.1 Flash-Lite	$0.25	$1.50	$0.025	none
Gemini 2.5 Flash-Lite	$0.10	$0.40	$0.01	none (older generation)

Output is 5× input on Argon and Flash, 6× on Pro. That ratio matters later, when thinking tokens show up.

The example: a document extraction pipeline

Say you run a pipeline that reads inbound documents (contracts, invoices, support emails) and extracts structured fields. It isn't an agent: one request per document, no back-and-forth. These numbers are an illustration; swap in your own.

• Instructions, JSON schema and a few worked examples: 5,000 tokens, identical on every request
• The document itself: 3,000 tokens
• The extracted JSON: 400 tokens
• Volume: 10,000 documents a day, so 300,000 a month

That's 8,000 input tokens and 400 output tokens per request. A month is 2.4 billion input tokens and 120 million output tokens.

To get your own per-request numbers, paste a real prompt into a token counter that counts for Gemini, Claude and GPT. Gemini averages about 4 characters per token in English, so a 3,000-token document is roughly 12,000 characters.

What does it cost on each Gemini model?

Model	Input	Output	Monthly today	After the increase
Gemini 4 Argon	$4,800	$1,200	$6,000	$12,000 (Dec 1)
Gemini 3.1 Pro Preview	$4,800	$1,440	$6,240	no change
Gemini 3.8 Flash	$1,800	$450	$2,250	$4,500 (Jan 1)
Gemini 3.5 Flash-Lite	$720	$300	$1,020	no change
Gemini 3.1 Flash-Lite	$600	$180	$780	no change
Gemini 2.5 Flash-Lite	$240	$48	$288	no change

Three things stand out.

Input is 77–80% of the bill on the bigger models. The 400-token answer is the cheap part. The 5,000 tokens of instructions you send 300,000 times a month are not.

The ranking flips on December 1. Today Argon is slightly cheaper than Gemini 3.1 Pro for this job ($6,000 vs $6,240). From December 1 it costs almost twice as much ($12,000 vs $6,240). If 3.1 Pro passes your evals, it's the cheaper strong model from then on. It's still labelled Preview, so weigh that too.

Flash-Lite is 2–8× cheaper than Flash. Extraction into a fixed schema is the kind of task small models often handle. Run your eval set on 3.1 Flash-Lite before assuming you need Flash.

I checked the table with this Gemini API cost calculator, entering 8,000 input and 400 output tokens at 300,000 requests a month. It also shows the post-increase prices.

Why is my Gemini bill higher than the math?

Usually because of one of three things the table above doesn't include.

1. Thinking tokens

Gemini's newer models think before they answer, and thinking tokens are billed as output at the full output price. They don't show up in the response text, so you only see them in the usage metadata.

Say Gemini 3.8 Flash thinks for 1,500 tokens before writing the 400-token JSON. Output per request goes from 400 to 1,900 tokens:

Gemini 3.8 Flash	Output tokens / month	Output cost	Monthly total
No thinking	120M	$450	$2,250
1,500 thinking tokens	570M	$2,137.50	$3,937.50 (+75%)
Same, from Jan 1	570M	$4,275	$7,875

Field extraction rarely needs long reasoning. Set a low thinking budget (or thinking level, depending on the model) in the request config, and compare accuracy on your eval set with thinking low and off.

2. Grounding with Google Search

Grounding is billed per search, not per token. Gemini 3 and newer models share 5,000 free search requests a month, then $14 per 1,000. If every one of the 300,000 documents triggered one search, that's (300,000 − 5,000) × $14 / 1,000 = $4,130 a month, more than the whole Flash bill. A grounded request can trigger more than one search, so treat that as a floor. Turn grounding on only for the requests that need fresh facts.

3. Long prompts on Pro

Gemini 3.1 Pro bills the whole request at $4 input and $18 output once the prompt passes 200,000 tokens. This pipeline never gets close, but a "just send the whole repo" coding tool or a long chat history does. Gemini 2.5 Pro has the same threshold ($1.25 → $2.50 input).

Does caching lower Gemini API cost?

Yes, and this pipeline is close to the ideal case: the same 5,000-token prefix starts every request. Cached input costs 10% of the input price.

Gemini has two kinds of context caching:

• Implicit caching is automatic. When a request starts with the same tokens as a recent one, those tokens are billed at the cached price. Nothing to set up, no storage fee, but hits aren't guaranteed. Put the fixed instructions first and the document last, or nothing matches.
• Explicit caching stores a prefix you create, for a time you choose, with guaranteed hits and a storage fee per million tokens per hour: $0.50 on Gemini 3.8 Flash. Gemini 4 Argon supports implicit caching only.

The prefix has to be long enough to cache: 4,096 tokens on these models. A 2,000-token system prompt doesn't qualify, which is one reason to keep your worked examples in the prefix rather than trimming them to save tokens.

If all 5,000 prefix tokens hit the cache:

Model	Without caching	With the prefix cached	Saved
Gemini 4 Argon (today)	$6,000	$3,150	47.5%
Gemini 4 Argon (from Dec 1)	$12,000	$6,300	47.5%
Gemini 3.8 Flash (today)	$2,250	$1,237.50	45%

On Flash, an explicit cache for that prefix costs 0.005M tokens × $0.50 × 720 hours = $1.80 a month to keep warm. For a pipeline that runs all day, buying guaranteed hits is an easy call.

Does the Batch API cancel the price increase?

For this pipeline, exactly. The Batch API takes 50% off input and output for jobs submitted asynchronously, and the increase is 2×:

Model	Standard, after increase	Batch, after increase
Gemini 4 Argon	$12,000	$6,000
Gemini 3.8 Flash	$4,500	$2,250

The catch is latency. Google targets a 24-hour turnaround (often much sooner), so batch only fits work nobody is waiting on: nightly extraction, backfills, evals, classification. If documents land in a queue anyway, a few hours' delay is often invisible to users. Batch also stacks with caching.

Is the Gemini API free?

Partly. Gemini 3.8 Flash, the Flash-Lite models and the Gemini 2.5 models have a free tier in Google AI Studio and the API, within rate limits. Gemini 4 Argon and Gemini 3.1 Pro are paid only.

The catch is data: on the free tier, Google may use your prompts and responses to improve its products. On the paid tier it doesn't. The free tier is fine for prototyping. For customer documents like the ones in this example, you want the paid tier.

Price your own requests

Every response carries usage_metadata. Log it, then price it. With the Python google-genai SDK:

PRICES = {  # USD per 1M tokens, Gemini 3.8 Flash until Dec 31, 2026
    "input": 0.75, "output": 3.75, "cache_read": 0.075,
}

def request_cost(usage) -> float:
    cached = usage.cached_content_token_count or 0
    fresh = usage.prompt_token_count - cached        # prompt count includes cached tokens
    output = usage.candidates_token_count + (usage.thoughts_token_count or 0)
    return (
        fresh * PRICES["input"]
        + cached * PRICES["cache_read"]
        + output * PRICES["output"]
    ) / 1_000_000

Two fields catch people out: thoughts_token_count is billed as output even though it isn't in the answer, and cached_content_token_count tells you whether implicit caching is actually hitting. If it's zero on the second request with the same prefix, your prefix isn't stable.

What to do before December 1

1. Log usage_metadata for a day of real traffic. Get average input, output, thinking and cached tokens per request.
2. Reprice it at the new prices: $4 / $20 for Argon, $1.50 / $7.50 for 3.8 Flash. Put that number in next quarter's budget, not today's.
3. If you're on Argon, run your eval set on Gemini 3.1 Pro and on a Flash model. From December 1, Pro costs half of Argon's input price and 60% of its output price.
4. Move anything nobody waits on to the Batch API. For async work, that undoes the increase.
5. Put the fixed instructions first, check that cached tokens show up from the second request, and set a thinking budget on simple tasks.

Are you moving off Argon before December 1, or absorbing the increase? I'm curious what your evals say about 3.1 Pro versus Argon on the same job.