Last updated: October 13, 2026

Gemini 3.8 Flash: API guide, pricing and real-world tests

Google released Gemini 3.8 Flash on September 2, 2026, three weeks after 3.7 Flash, together with a restricted security model, 3.8 Flash Cyber. Flash is Google’s fast, lower-cost tier, now pitched at agent workflows. This page shows how to call it, what it costs now and from January 2027, how to use it for free, and how it compares with Claude Haiku 5.5 and Sonnet 5.5.

Gemini 3.8 Flash at a glance

Google has not published a knowledge cutoff for 3.8 Flash on the model page.

Source: Gemini API model page

Source: Google launch post

Gemini 3.8 Flash pricing

Google’s launch price is half the standard price and runs through December 31, 2026. One input price covers text, images, video and audio, and thinking tokens are billed as output.

• There is no long-context surcharge: the same price applies up to the full 1M-token window.
• Grounding with Google Search: 5,000 requests a month free (shared across Gemini 3 and newer models), then $14 per 1,000.
• Gemini 3.7 Flash is still supported for cost-first workloads, but Google’s pricing page no longer lists a price for it.

Source: Gemini API pricing

Source: Google launch post

How to use Gemini 3.8 Flash for free

1. Open Google AI Studio and sign in with a Google account.
2. Pick Gemini 3.8 Flash in the model list and chat with it, or click Get API key.
3. Use the key without a billing account: the free tier covers 3.8 Flash input and output.
4. Check your free rate limits in AI Studio. Google does not publish them on the docs page.

On the free tier Google may use your prompts and responses to improve its products; on the paid tier it does not. Do not send private code or customer data through a free key. In the Gemini app, 3.8 Flash is listed for Google AI Pro and Ultra subscribers.

Source: Google AI Studio

Source: Gemini API pricing

API guide: call Gemini 3.8 Flash from Python

1. Create a key in Google AI Studio and save it as GEMINI_API_KEY.
2. Install the Google Gen AI SDK: pip install -U google-genai.
3. Call the Interactions API with model gemini-3.8-flash.
4. Start with thinking_level low for chat and extraction; raise it to medium or high for multi-step coding or agent tasks.

# pip install -U google-genai
from google import genai

client = genai.Client()  # reads GEMINI_API_KEY

interaction = client.interactions.create(
    model="gemini-3.8-flash",
    input="Explain what a race condition is, with a short Python example.",
    generation_config={"thinking_level": "low"},  # low | medium | high
)
print(interaction.output_text)

Source: Gemini API text generation docs

REST call with curl

curl -X POST "https://generativelanguage.googleapis.com/v1beta/interactions" \
  -H "x-goog-api-key: $GEMINI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gemini-3.8-flash",
    "input": "How does AI work?"
  }'

Source: Gemini API text generation docs

OpenAI-compatible endpoint

Tools built on the OpenAI SDK only need the base URL, key and model name changed. reasoning_effort maps to Gemini’s thinking levels; do not send it together with thinking_level.

# pip install openai
from openai import OpenAI

client = OpenAI(
    api_key="GEMINI_API_KEY",
    base_url="https://generativelanguage.googleapis.com/v1beta/openai/",
)

resp = client.chat.completions.create(
    model="gemini-3.8-flash",
    reasoning_effort="low",  # maps to thinking level low / medium / high
    messages=[{"role": "user", "content": "Summarize this changelog in 5 bullets: ..."}],
)
print(resp.choices[0].message.content)

Source: OpenAI compatibility docs

Benchmarks Google published

These are Google’s own numbers, with the comparison models exactly as Google labelled them (older Claude and GPT versions, not Claude 5.5). Google gives no test settings or run counts.

• DeepSWE v1.1 (long-horizon software engineering): over 70% success, shown on a chart without exact figures for the other models.
• Glean, a launch customer, says 3.8 Flash completed more than three times as many tasks as 3.7 Flash in its own evaluations.
• Google says 3.8 Flash uses more reasoning steps and tool calls on hard tasks, so it can spend more tokens at higher effort levels.

Source: Google DeepMind Flash page

Source: Google launch post

Independent test: Artificial Analysis

Artificial Analysis runs the same suite of tasks on every model, which makes it the closest thing to an apples-to-apples comparison. Figures read on 2026-10-13.

Haiku 5.5 scores slightly higher than 3.8 Flash on this index and costs about a sixth as much per task, while Flash uses far fewer tokens. Sonnet 5.5 is clearly stronger and more expensive.

Source: Artificial Analysis: Gemini 3.8 Flash

Source: Artificial Analysis: Claude Haiku 5.5

Source: Artificial Analysis: Claude Sonnet 5.5

Real use cases: what to test it on

We have not run 3.8 Flash ourselves. The quickest way to judge it is to run the same prompts against it and the model you use now:

• Agent workflows. Google’s main pitch. Finance and legal agent benchmarks are where its published lead is clearest.
• Video, audio and PDF input. One model and one price for all input types. Claude Haiku 5.5 and Sonnet 5.5 take text and images only.
• Search-grounded answers. Built-in Google Search grounding with a monthly free allowance, useful for news or product lookups.
• Bulk jobs. Batch and Flex halve the price to $0.375 / $1.875 for work that can wait.

Watch this screen recording. List the exact steps that reproduce the bug, what the user expected, and which UI component is most likely at fault.

From the attached invoice PDFs, return JSON with vendor, invoice number, date, currency, line items and total. Use null for anything missing.

Using Google Search, find this week’s pricing changes for the three biggest cloud GPU providers. Give one line per change with the source URL.

You have shell access to a Python repo. Plan how you would upgrade it from Pydantic v1 to v2 and how you will check nothing broke. Do not run anything yet.

Source: Google launch post

Source: Gemini API model page

Gemini 3.8 Flash vs Claude Haiku 5.5 and Sonnet 5.5

Haiku 5.5 vs Gemini 3.8 Flash · Sonnet 5.5 vs Gemini 3.8 Flash · All model prices

Source: Gemini API pricing

Source: Anthropic pricing

What 1,000 requests cost

List prices, no caching or batch discounts. Real bills also depend on each model’s tokenizer and on how many thinking tokens it spends.

Source: Gemini API pricing

Source: Anthropic pricing

Gemini 3.8 Flash Cyber

A variant with looser cybersecurity safeguards for defenders. It is not on the public API: access is through Google’s Fairwind Program, with priority for government agencies, critical infrastructure operators and software maintainers. Google reports 47.2% pass@1 on CWE-Bench patching (vs 47.8% for a leading frontier model) and 2.6x more correct Chrome vulnerability patches than the best commercial models.

Source: Google launch post

Source: Fairwind Program

Gemini 3.8 Flash TTS

Text-to-speech is a separate model, gemini-3.8-flash-tts, released on September 23, 2026. It costs about $0.81 per hour of audio at the launch price. See the full TTS guide for code, voices and pricing.

Gemini 3.8 Flash TTS guide

Source: Google TTS launch post

Independent guide, not affiliated with Google or Anthropic. Prices and benchmarks change; check the linked official pages before you commit to a budget.