Gemini models in the API: 3.8 Flash vs 3.7 Flash and which to pick

If you are searching "gemini api models" and want to know which Gemini model to call, the Zukko catalogue on October 18, 2026 has two: gemini-3.8-flash and gemini-3.7-flash. Their price is the same, the context is the same; the strength index and age differ. Below we compare them by the numbers, explain what a 1M-token window means for cost, price a batch job and show how to get the current model list from the API so your code does not break when versions change.

Gemini API models: two models in the catalogue

Data from the Zukko table (prices in dollars per 1M tokens, on October 18, 2026):

The official price row is there so you can check the calculation: at Zukko input on gemini-3.8-flash costs $0.125 against $0.75 from the provider, output $0.6 against $3.75. The current list and prices: GET https://api.zukko.pro/v1/models and the "Models" page in the dashboard.

How 3.8 Flash differs from 3.7 Flash

In price and context, not at all. That is an important conclusion: the choice between them does not affect the bill, so you have to decide on other grounds.

Strength index. This is the Intelligence Index from Artificial Analysis: 41 for gemini-3.8-flash against 38 for gemini-3.7-flash. The index is a general guide, not a quality score on your task. A three-point gap does not guarantee that 3.8 will handle your prompt better. Test it.

Age. Older models are eventually retired by providers. When a model is retired or switched off, requests to it start returning an error, for example 503 model_disabled. Code hard-wired to gemini-3.7-flash will break one day. That is an argument for defaulting to the newer version.

Reproducibility. The only sensible reason to keep 3.7 is a finished pipeline whose results you have already verified and do not want to change without running tests. Then do the move to 3.8 as a separate task: run a set of examples, compare answers, then switch.

Images. Both models are marked "image input: no" in the table. If you need to send screenshots or scans, choose a model marked "yes", for example gpt-6.1-sol, claude-sonnet-5-5 or qwen3.8-max.

The Gemini context window and cost

Both models have a 1M-token window. It is the maximum per request, and you pay for every token you send. Example: a request that fills the window completely costs, on gemini-3.8-flash, 1,000,000 × 0.125 / 1,000,000 = $0.125 for input, before any answer. If you repeat such a request ten times with different questions about the same document, input comes to $1.25, and with cache hits (read $0.013) noticeably less. More on what a long context costs is in the article on long context and price↗.

How many tokens your document takes depends on the language: Russian text usually takes more tokens than English of the same length. Do not plan by word count, measure on your own texts.

A long window does not mean you should always fill it. For a question about one page of a document it is cheaper to pass that page rather than the whole file.

Cache on Gemini

Cache read costs $0.013 per 1M tokens against an input price of $0.125. The rule is the same as for models other than Claude and GPT: cache reads are counted if the provider reports them in the response, and a separate cache write is not counted. You will see what actually hit the cache in the bill under "Usage", so do not bake the discount in beforehand but measure it on your own request.

Pricing a batch job

The task: process 10,000 documents, each with 3,000 input tokens and 300 output tokens. In total input is 30,000,000 and output 3,000,000.

• Zukko, gemini-3.8-flash or gemini-3.7-flash: 30 × 0.125 + 3 × 0.6 = 3.75 + 1.80 = $5.55.
• At official prices ($0.75 / $3.75): 30 × 0.75 + 3 × 3.75 = 22.50 + 11.25 = $33.75.
• If 2,000 of the 3,000 tokens in each request (a shared instruction) are read from cache: fresh input is 10,000 × 1,000 = 10,000,000 tokens, cache 20,000,000. Cost = 10 × 0.125 + 20 × 0.013 + 3 × 0.6 = 1.25 + 0.26 + 1.80 = $3.31. This assumes the provider returns cache information.

By price, Flash models are aimed at mass work: classification, extraction, short answers. For such processing build in retries on temporary errors and record usage for every request. For tasks that need high strength, compare with higher-tier models, for example claude-opus-5-5 (index 58): the price there is higher, and whether it pays off is shown only by a test on your own data.

How to get the current list through /v1/models

The model list changes. This endpoint is public, no key needed, and the request is free:

curl -s https://api.zukko.pro/v1/models | jq -r '.data[].id' | grep -i gemini

If new ids appear in the output, the catalogue has been updated. Code should not assume gemini-3.7-flash will live forever. The minimum protection against model changes:

import os

from openai import APIStatusError, OpenAI

client = OpenAI(
    base_url="https://api.zukko.pro/v1",
    api_key=os.environ["ZUKKO_API_KEY"],
)

MODELS = ["gemini-3.8-flash", "gemini-3.7-flash"]  # in order of preference
SWITCH_CODES = {"model_retired", "model_disabled", "model_not_found"}

def ask(prompt: str):
    last_error = None

The function tries models in order and moves to the next only if the model is retired, disabled or not found. Other errors (401, 402, 429) are re-raised because changing the model will not fix them. The topic is covered in detail in how to get the model list and not break when it changes↗; the endpoint itself is described in the Zukko docs, section GET /v1/models.

Which model to take

• A new project: gemini-3.8-flash. Same price, higher index, and the provider will retire the older version first.
• A working pipeline on 3.7: keep it until you have run your checks on 3.8.
• You need image input: neither one; look at the models marked "yes" in the table.
• You need maximum cheapness: compare with deepseek-v4.1-flash ($0.013 / $0.038, index 39) and glm-5.3-flash ($0.04 / $0.04, index 42, context 128K). Cheaper does not mean suitable: test on your own tasks.

A detailed comparison of Gemini Flash with DeepSeek is in the article DeepSeek or Gemini Flash↗.

FAQ

Do I need a separate Google key for the Gemini API?

No. Through Zukko the call goes with a fetch_... key in the OpenAI-compatible format at https://api.zukko.pro/v1, and you put the model name in the model field. The key is created in the dashboard↗.

Can I call Gemini in the Anthropic format?

Any model can be called in either format, OpenAI or Anthropic. For Gemini /v1/chat/completions is the more usual choice.

What is Gemini's context window?

Both models in the Zukko catalogue have 1M tokens. For long documents estimate the real size of your texts in tokens and remember that you pay for every token you send.

How do I pay from Russia?

Gemini is called through the gateway with a fetch_... key, and the balance is topped up via SBP, a MasterCard card or crypto. More: Gemini in Russia through the API↗.