Google released Gemini 3.8 Flash on September 2 under the stable model ID gemini-3.8-flash, and the Gemini API lists it as generally available. This is not a context-window upgrade: the model still accepts up to 1,048,576 input tokens and produces up to 65,536 output tokens. The important change is its role. Google is positioning Flash not merely as a cheap, fast answer engine, but as a model for long-horizon software engineering, autonomous agents, and complex knowledge work.

That creates the central tension in this release. Token prices match 3.7 Flash and output throughput is faster, yet 3.8 may reason longer, invoke more tools, and verify more work. A completed task is not guaranteed to cost the same.

What the model actually supports

Gemini 3.8 Flash accepts text, images, video, audio, and PDFs, but its output is text only. It supports function calling, structured output, code execution, file search, URL Context, Google Search and Maps grounding, context caching, and Batch, Flex, and Priority inference. Computer Use is available as a preview capability.

The exclusions matter. It does not generate images or audio, does not support the Live API, and is not available for tuning. “Multimodal” therefore describes understanding several input formats, not a single model that creates every media type. Developers can use it through AI Studio and the Gemini API. Google also lists it across the Gemini app, Search AI Mode, and Sheets for Pro and Ultra subscribers, rather than promising identical access to every free account and region.

A Flash model that may think before it speaks

The API exposes low, medium, and high thinking levels, with medium as the default. minimal is not supported and returns an error. Low is a sensible starting point for chat, classification, drafting, and latency-sensitive tool calls. Medium targets most coding and agent workflows. High gives difficult reasoning and long tool loops more compute, but can increase thinking tokens, tool turns, and time to first token.

Google describes 3.8 as more diligent on hard tasks. Independent measurements from Artificial Analysis point in the same direction: at high effort, 3.8 scores above 3.7 and has higher steady-state output speed, but it takes longer to deliver the first token and its average task cost is roughly 40 percent higher. Its advantage is less “every answer appears instantly” and more “the model keeps working and checking.”

The largest gains are in execution

Google’s strongest reported improvements appear in terminal work, multi-file software engineering, finance agents, and other tool-driven tasks. Several general knowledge and multimodal results move only slightly, and the ordinary Humanity’s Last Exam result is marginally below 3.7. Vendor tables are useful evidence of optimization direction, but they do not prove universal superiority: model providers may use different agent harnesses, thinking settings, or self-reported competitor scores.

Independent evaluation shows a similar tradeoff. Higher completion scores arrive with more generated tokens and more agent steps. A production team should therefore compare end-to-end task success, retries, tool charges, elapsed time, and human repair—not just a price per million tokens or one leaderboard percentage.

The list price and the task bill are different

The introductory standard API rate is $0.75 per million input tokens and $3.75 per million output tokens, including thinking tokens. Google has already announced that the later standard rate will double. Batch and Flex inference cost less for work that does not need an immediate response. A free developer tier exists, but it has quotas and different data-use terms; it is not unlimited free inference.

For document-heavy and agentic systems, three controls matter immediately. Cache repeated context, route thinking level by task difficulty, and record total input, output, tool activity, and successful-task cost. High effort can be economical if it prevents failed runs and human rework. Applying it to simple extraction or classification merely buys extra deliberation.

Start with the Python SDK

Install the current Google GenAI SDK and supply the API key through the normal environment configuration:

pip install -U google-genai

from google import genai
from google.genai import types

client = genai.Client()

response = client.models.generate_content(
    model="gemini-3.8-flash",
    contents="Review this retry workflow and identify race conditions.",
    config=types.GenerateContentConfig(
        thinking_config=types.ThinkingConfig(
            thinking_level="medium"
        )
    ),
)

print(response.text)

Migration from 3.7 or 3.6 deserves more than a model-name replacement. Convert an integer thinking_budget to thinking_level, then regression-test schemas, function arguments, tool ordering, timeouts, and cost. Google’s developer guide also removes or rejects several legacy sampling and penalty parameters. Validate the request shape during deployment instead of relying on accidental backend compatibility.

Flash Cyber is not another public toggle

Google released Gemini 3.8 Flash Cyber alongside the standard model. They share foundational intelligence, but Cyber is specialized for vulnerability discovery and automated patching and is distributed only to trusted governments, critical-infrastructure operators, and software maintainers through the Fairwind Program. Its security mitigations are intentionally more permissive. Cyber benchmark results should not be attributed to the public gemini-3.8-flash API.

My assessment

The most important signal is that Google is redefining Flash from a high-throughput small-model tier into an agent execution layer. A stable API, broad tool support, and a low entry price make 3.8 a strong candidate for A/B tests in coding, retrieval, and multi-step workflows. They do not justify an immediate production-wide replacement based on a chart.

The model card still acknowledges hallucinations, occasional slowness, and timeouts. Higher effort can use substantially more tokens, and Google’s internal multilingual safety metric shows a regression. The practical adoption path is to run 3.7 and 3.8 on the same real tasks, measure successful-task cost and human correction, and route only the workloads that benefit. Gemini 3.8 Flash is not a discounted Pro model; it is a fast agent workhorse that is increasingly willing to run a few extra laps.