Gemini 3.8 Flash pricing: intro rates, thinking tokens, and the real cost per task

Gemini 3.8 Flash costs $0.75 per million input tokens and $3.75 per million output tokens, the same introductory rate Google set for 3.7 Flash. That rate runs through December 31, 2026. On January 1, 2027 it doubles to $1.50 and $7.50, and the same doubling hits 3.6 Flash and 3.7 Flash on the same day. Nothing about the per-token price changed with this release.

What did change is how many tokens the model spends. Google says 3.8 Flash “works harder”: it takes more reasoning steps, calls tools iteratively, and “can use more tokens on longer running and complex tasks, by design”. Artificial Analysis measured the effect independently: running its Intelligence Index cost $0.58 per task on 3.8 Flash at high thinking versus $0.40 on 3.7 Flash, with about 30% more output tokens per task. Same sticker price, higher bill.

This guide walks every line on the official Gemini API pricing page, works a 1,000-tasks-a-day example at each thinking level, and compares the rate against Flash-Lite, Claude Sonnet 5, and GPT-5.6 Luna. For the model overview first, start with what Gemini 3.8 Flash is.

Gemini 3.8 Flash pricing at a glance

All prices are per 1 million tokens on the paid tier.

Three details matter more than the headline numbers. Output pricing covers thinking tokens, so internal reasoning bills at $3.75, not $0.75. The intro rate has a hard end date. And the 3.7 Flash and 3.6 Flash rows carry the identical Jan 1 doubling, so an older Flash does not shield you from the increase. The 3.7 Flash pricing breakdown covers the same rates on a model that spends fewer tokens per task.

Thinking tokens are output tokens

Gemini 3.8 Flash reasons before it answers, and every token of that reasoning is metered as output. The API reports the count as usageMetadata.thoughtsTokenCount on a generateContent response, separate from candidatesTokenCount (the visible answer) and promptTokenCount (your input). Both output buckets bill at the same $3.75.

You steer the amount with thinking_level: low, medium, or high. On 3.8 Flash the default is medium, not high as on Gemini 3 Pro. minimal is gone and returns a validation error, so any config carried over from earlier models needs that value mapped to low. Google’s thinking documentation describes the levels as relative effort, not a token budget, so you cannot cap thinking tokens directly the way the old integer thinking_budget allowed. What each level does to latency and cost is in the 3.8 Flash thinking levels guide.

A quick illustration with made-up but realistic sizes. Suppose a request sends 10,000 input tokens and gets back 2,000 visible output tokens plus 6,000 thinking tokens. At intro rates:

• Input: 10,000 x $0.75 / 1,000,000 = $0.0075
• Output: (2,000 + 6,000) x $3.75 / 1,000,000 = $0.03
• Total: $0.0375 per request

Thinking is 75% of the output charge and 60% of the whole request. From January 1 the same request costs $0.015 + $0.06 = $0.075. “Same price as 3.7” is true and incomplete: the rate held, the token count moved.

Why the cost per task went up while the price didn’t

Google’s launch post is direct about the trade: on complex tasks the model “executes extra reasoning steps, and calls tools iteratively”, verifying its work along the way. More steps means more thinking tokens and, in agent loops, more tool-call turns. Google’s mitigation advice: lower thinking_level, or stay on 3.7 Flash.

Artificial Analysis put numbers on it. Its Intelligence Index runs nine evals; 3.8 Flash at high scored 59 against 56 for 3.7 Flash at high, using about 48,000 output tokens per task on average, up 30% from 3.7 Flash. The cost to run the index per task:

Read the table two ways. Against its predecessor, 3.8 Flash at high costs 45% more per task ($0.58 / $0.40 = 1.45) for a three-point index gain. Against itself, dropping from high to medium cuts the per-task cost by 29% ($0.41 / $0.58 = 0.71), and low cuts it by 59% ($0.24 / $0.58 = 0.41). Medium on 3.8 Flash lands within a cent of high on 3.7 Flash, a useful anchor when deciding whether to upgrade at all. The 3.8 Flash vs 3.7 Flash comparison works through that decision by workload.

Artificial Analysis also lists a blended price of $0.58 per million tokens at a 3:1 input-to-output ratio. That it matches the high-level cost per task is a coincidence; one is a per-token rate, the other depends on how many tokens the model chooses to spend.

Lock in the intro rate before December 31

“Lock in” needs a precise meaning, because the intro rate is not a contract. Google bills usage at the rate in force when tokens are consumed, so you cannot prepay 2027 usage at 2026 prices, and moving to 3.7 or 3.6 Flash does not help. What you can do is shift discretionary work into the window:

• Run backfills and one-off document processing now through the Batch API. A 100 million token backfill (75M input, 25M output) costs 75 x $0.375 + 25 x $1.875 = $28.13 + $46.88 = $74.99 in batch this year, and twice that, $149.98, in January.
• Build your eval set and baseline token counts before the price changes, so the January bill has one variable moving, not two.
• Decide the thinking level per route this quarter. Every route left on medium by default is a route whose cost you have not examined. Routes that pass evals at low should be on low before January.

If price is the deciding factor across providers, our cheapest LLM API providers roundup lays out the full field; the Fable 5.1 and GPT-5.6 Sol comparison covers the premium tiers.

How it compares to Flash-Lite, Sonnet 5, and GPT-5.6 Luna

Per-token rates, per 1 million tokens:

At intro rates, 3.8 Flash input is 2.5x Flash-Lite’s and output is 1.5x. Against Sonnet 5, 3.8 Flash is about 2.7x cheaper on both input ($2 / $0.75) and output ($10 / $3.75). Against Luna it is cheaper too: $0.75 vs $1 on input, $3.75 vs $6 on output.

The picture flips on January 1. At $1.50 and $7.50, 3.8 Flash becomes more expensive than Luna on both sides, and the gap to Sonnet 5 narrows to about 1.3x ($2 / $1.50 and $10 / $7.50). Flash-Lite stays cheaper throughout. If the intro rate is why you chose 3.8 Flash, put the January re-evaluation on the calendar now.

Per-token comparison is only half of it. A model that spends fewer tokens per answer can cost less per task at a higher rate, and the only way to know is to run the same task set through each candidate and read the usage counts back. The 3.8 Flash API guide shows how to read usageMetadata from both the Interactions API and legacy generateContent.

Catch cost regressions with Apidog token assertions

The failure mode with 3.8 Flash is not a wrong answer. It is a right answer that quietly used 40% more tokens after a prompt tweak or a thinking-level change, and nobody notices until the invoice. Apidog fixes that by treating the token count as a field to assert on, like a status code.

1. Store the key as an environment variable. Create GEMINI_API_KEY in an Apidog environment and reference it as {{GEMINI_API_KEY}} in the x-goog-api-key header, so the key never sits in a saved request.
2. Send a golden prompt. Save a POST to https://generativelanguage.googleapis.com/v1beta/models/gemini-3.8-flash:generateContent with a representative prompt and an explicit thinkingConfig.thinkingLevel, one per production route.
3. Assert on the usage fields. Add a post-processor script that reads usageMetadata and fails the test if tokens exceed a ceiling set from your baseline:

const usage = pm.response.json().usageMetadata;
pm.test("thinking tokens within budget", () => {
  pm.expect(usage.thoughtsTokenCount).to.be.below(8000);
});
pm.test("visible output within budget", () => {
  pm.expect(usage.candidatesTokenCount).to.be.below(2500);
});
pm.test("request cost within budget", () => {
  const cost = usage.promptTokenCount * 0.75 / 1e6
    + (usage.thoughtsTokenCount + usage.candidatesTokenCount) * 3.75 / 1e6;
  pm.expect(cost).to.be.below(0.05);
});

4. Schedule it. Put the requests in a test scenario and run it on a schedule, so a token-usage jump shows up as a failed run the same day; our guide to scheduling API tests in Apidog covers the setup. Update the two rate constants on January 1 and the cost assertion keeps tracking the bill.

The same scenario doubles as a provider bake-off: point a copy at Flash-Lite, Sonnet 5, or Luna and compare the usage fields side by side. Download Apidog to build the first scenario; the free plan covers this workflow.

FAQ

Does Gemini 3.8 Flash cost more than 3.7 Flash? Per token, no. Both are $0.75 input and $3.75 output through December 31, then $1.50 and $7.50. Per task, yes: Artificial Analysis measured $0.58 per index task on 3.8 Flash at high versus $0.40 on 3.7 Flash, because 3.8 Flash generates about 30% more output tokens.

Are thinking tokens billed separately? No. They bill as output tokens at $3.75 per million (intro) and show up in usageMetadata.thoughtsTokenCount. You control them indirectly through thinking_level; there is no hard token budget on 3.8 Flash, and minimal returns an error.

Is there a free tier for Gemini 3.8 Flash? Yes. The AI Studio free tier charges nothing for input or output, with rate limits shown in AI Studio, and Google uses free-tier data to improve its products. The consumer Gemini app serves 3.8 Flash only to AI Pro and Ultra subscribers. Details are in the free-usage guide.

What happens to my costs on January 1, 2027? Input, output, cache read, cache storage, and batch rates all double. If token usage per task stays the same, your bill doubles too. The 3.6 and 3.7 Flash rows double on the same date.

Which thinking level keeps costs down? Start from Artificial Analysis’s spread: low $0.24, medium $0.41, high $0.58 per index task. Run your own evals at each level, keep the lowest one that passes, and assert on token counts so the setting cannot drift.

What to do this week

Gemini 3.8 Flash is priced like 3.7 Flash and spends tokens like a bigger model. Treat the thinking level as a cost setting and set it per route. Move batch-friendly work into the intro window before December 31. Baseline your token usage now and assert on it, so the January doubling is a change you priced in.