Gemini 3.8 Flash API: Should You Upgrade? Pricing and Benchmarks

Official Gemini 3.8 Flash and Gemini 3.8 Flash Cyber launch artwork. Source: Google's announcement.

The Gemini 3.8 Flash API launched on September 2, 2026 with the stable model id gemini-3.8-flash. It accepts up to 1,048,576 input tokens, handles text, images, video, audio and PDFs, and can work with function calling, code execution, search, URL Context and Computer Use.

The release is less about one benchmark win than a change in how the Flash tier handles long jobs. Google positions Gemini 3.8 Flash for production agents, coding and knowledge work. On hard tasks, it can take shorter reasoning steps and make more tool calls than Gemini 3.7 Flash. That can improve progress through a workflow, but it can also produce more billable tokens.

The launch price has a firm end date. Global Standard API usage costs $0.75 per million input tokens and $3.75 per million output tokens through December 31, 2026. On January 1, 2027, those rates become $1.50 and $7.50. This guide separates model facts, Google-reported evaluations, official UI demos and migration risks so you can decide what to test.

Gemini 3.8 Flash API at a glance

The official Gemini 3.8 Flash model page is the source of truth for formats and capabilities. Multimodal input does not mean multimodal output: 3.8 Flash can inspect media, but its native response is text. Use a separate image model if the workflow must return an image.

A one-million-token window can hold a large repository, a long recording or a document set. It should not remove retrieval, caching or trimming from the architecture. Long prompts raise latency and cost, and irrelevant context can make tool decisions worse. The larger window gives an orchestrator room to choose; it does not make every available token useful.

What is in the Gemini 3.8 Flash family?

Google announced two models with similar names and different access rules.

Gemini 3.8 Flash is the general model covered in this guide. Developers can use it through the Gemini API, Google AI Studio, Vertex AI and Gemini Enterprise Agent Platform. Google AI Pro and Ultra subscribers also get access through the Gemini app, AI Mode and Google Sheets.

Gemini 3.8 Flash Cyber is built for cyber defense. Access runs through Google DeepMind's Fairwind program for approved, trusted defenders working on areas such as vulnerability analysis and incident response. It is not a public model id that any API customer can select, and approved organizations face added use and redistribution conditions.

Most product teams should plan around gemini-3.8-flash. Flash Cyber is relevant only if the organization qualifies for and joins Fairwind.

Gemini 3.8 Flash API pricing

Google currently lists promotional Global Standard API rates and the prices that take effect after the promotion. Each number below is per one million tokens.

Suppose one repository task sends one million uncached input tokens and generates 200,000 output tokens. The model charge during the promotion is about $1.50: $0.75 for input and $0.75 for output. The same token mix costs about $3.00 at the regular Standard rate. Search, Maps and other tool charges are not included.

For offline work, the Gemini Batch API costs 50% of the Standard rate. Google targets completion within 24 hours and says jobs often finish sooner. Batch is a good fit for evaluation runs, bulk summaries and other asynchronous processing. It is not a substitute for an interactive chat or agent loop.

There are two pricing boundaries to keep in mind. Vertex AI rates can differ outside the global region or under service tiers such as Priority. Also, Gemini 3.8 Flash may generate more tokens than 3.7 on difficult tasks. Equal per-token pricing does not guarantee an equal invoice. Google calls out that behavior in its latest-model and migration guide.

What the Gemini 3.8 Flash benchmarks show

Google's launch table compares Gemini 3.8 Flash with Gemini 3.7 Flash, Claude Opus 5, GPT-5.6 Sol and other models across agent, coding and professional-work evaluations. Gemini 3.8 Flash scores 73.7 on DeepSWE v1.1 versus 65.3 for 3.7 Flash. It records 1545 on GDPVal-AA v2, 89.4 on Terminal-Bench 2.1 and 59.0 on OSWorld 2.0.

Google's launch table retains the original benchmark labels, footnotes and comparison models. Results come from vendors or the listed evaluation providers under specific settings; Mengbi did not reproduce them. Source: Google's Gemini 3.8 Flash launch post.

The results put 3.8 Flash in the leading group for long-running agent and coding work, but the table does not show a win in every row. Claude Opus 5 scores 74.0 on DeepSWE v1.1, just above 3.8 Flash at 73.7. Other rows use their own tools, prompts and time limits. The table is a useful shortlist, not a purchasing decision by itself.

DeepSWE v1.1 measures work on real software-engineering tasks. Gemini 3.8 Flash scores 73.7, close to the highest bar, but the agent runner and evaluation settings still shape the result. Source: Google's launch post.

A better internal test uses 30 to 100 representative tasks and records final test pass rate, human correction count, tool turns, input and output tokens, timeouts and full cost. A model that explores for many rounds can be more expensive in bulk even when its token rate looks low.

Four official Gemini 3.8 Flash UI demos

Google included four interactive projects in the launch post: a game, a retro map, a geospatial visualization and a 3D teaching tool. They show the sort of continuous front-end work 3.8 Flash can perform in Antigravity and Google AI Studio. They are selected showcases, not blind tests or evidence that each project worked on the first attempt.

Chronomancers: a game built through a looped instruction

Chronomancers was built in Google Antigravity through a looped instruction, with textures made using Nano Banana. Frame from Google's official demo video.

Chronomancers combines a world view, status panels, combat and generated assets in one playable interface. The useful unit of evaluation is the working game loop, not the appearance of one frame. A team testing the same pattern should score whether state, controls and recovery continue to work after several rounds.

DOS Google Maps: live mapping in a retro shell

DOS Google Maps was made in Antigravity. This frame preserves the code workspace, running interface and agent verification shown in Google's official video.

This project places live map behavior inside a DOS-style window. The demo exposes file structure, task progress, map controls and validation feedback. It is evidence that a model can work on visual and functional constraints in the same development loop. The result still depends on a map API, browser tools and the runtime around the model.

Topographic Map: live cross-sections from USGS data

Topographic Map uses USGS data for real-time cross-sections and projections. Frame from Google's official demo video.

The topographic project connects a 3D scene, source data, camera controls and a profile chart. It is closer to a small data product than a static landing page. Tests should cover coordinate edges, zoom behavior, missing data and lower-powered devices, not just the recorded path.

Hardware Anatomy: an explodable 3D learning tool

Hardware Anatomy was built with Three.js in Google AI Studio. A Deconstruct slider separates the parts. Frame from Google's official demo video.

Hardware Anatomy turns a Three.js scene into a teaching interaction. Dragging the slider exposes the assembly. A production version would still need keyboard control, text alternatives, mobile input and a fallback for weaker devices. Short launch videos rarely cover those requirements.

How to call the Gemini 3.8 Flash API

Create an API key in Google AI Studio and store it in the server-side environment variable GEMINI_API_KEY. Google's current Python SDK package is google-genai. Start with a plain request so authentication, the model id and the response shape can be checked before tools are added.

pip install -U google-genai

from google import genai
client = genai.Client()
response = client.models.generate_content(    model="gemini-3.8-flash",    contents="Review this deployment plan and return the three highest-risk assumptions.",)
print(response.text)

genai.Client() reads GEMINI_API_KEY. Do not place that key in a browser bundle, public repository or client-visible environment variable. A production service should also log request ids, token use, tool-call counts, latency and stop reasons, with limits for retries and total agent turns.

Use Gemini 3.8 Flash with the OpenAI Python SDK

Google exposes an OpenAI-compatible endpoint for projects that already use the OpenAI SDK. The compatibility layer remains beta, so run regression tests with the exact SDK version in your deployment.

import os
from openai import OpenAI
gemini_openai_host = "generativelanguage.googleapis.com"
client = OpenAI(    api_key=os.environ["GEMINI_API_KEY"],    base_url=f"https://{gemini_openai_host}/v1beta/openai/",)
response = client.chat.completions.create(    model="gemini-3.8-flash",    reasoning_effort="medium",    messages=[        {            "role": "user",            "content": "Create a testable migration checklist for this API change.",        }    ],)
print(response.choices[0].message.content)print(response.usage)

reasoning_effort accepts low, medium or high. Do not set the same thinking behavior through both a compatibility field and an overlapping native Gemini field. See Google's OpenAI compatibility guide for the full request surface.

Agents need more than a successful text response. The client must pass function declarations, return tool results and continue the model turn while enforcing timeouts, permissions, idempotency and cost limits.

Migrating from Gemini 3.7 Flash to 3.8

For a request that only sets model and contents, start by swapping the model id. Requests with sampling controls, thinking budgets, multiple candidates or tool turns need a closer pass through Google's migration guide.

Do not move all production traffic first. Begin with shadow requests or a 5% to 10% canary and compare 3.7 with 3.8 on completion rate, output length, tool errors and total spend. Expand only after rollback, timeout handling and usage logs have been checked.

The change in agent behavior can be easy to misread. A shorter reasoning step followed by another tool call may look responsive while adding turns to the job. Measure the cost and time needed to reach the accepted business result, not only latency for one response.

Gemini 3.8 Flash vs 3.7 Flash

Google has not announced a shutdown date for Gemini 3.7 Flash. Its deprecation page still lists the model as supported. Keep 3.7 as a rollback and efficiency control instead of retiring it only to follow the newest version.

If the model will sit inside an IDE or terminal workflow, see our AI coding agents comparison. For research, spreadsheets and broader operations, the AI agents for work guide maps the surrounding tools and human checks.

How to evaluate Gemini 3.8 Flash before launch

A reusable scorecard should record at least these fields:

1. Whether the task finished and the final test passed.
2. The number of follow-up prompts and manual edits.
3. Input, cached-input and output tokens, plus the full charge.
4. Total tool calls, failed calls and repeated calls.
5. P50 and P95 time to first token, response and completed task.
6. Timeouts, rate limits, empty responses and schema parse failures.
7. Safety refusals, factual errors and whether cited answers can be checked.

Do not build the set only from successful demos. Include ambiguous requests, dead links, oversized files, tool errors, duplicated function results and mobile screenshots. Keep the input, acceptance script and human scoring rule fixed for each item. Otherwise, the test compares prompts rather than model versions.

Google's model card lists hallucinations, occasional slowness or timeouts and higher token use at greater thinking effort among the known limitations. Not every safety evaluation improved over 3.7. Medical, legal, financial, privileged-action and cybersecurity workflows still require human review and system-level access controls.

Frequently Asked Questions

When was Gemini 3.8 Flash released?

Google released Gemini 3.8 Flash on September 2, 2026. Developers can access the stable gemini-3.8-flash model through the Gemini API, Google AI Studio and Vertex AI.

How much does the Gemini 3.8 Flash API cost?

Global Standard API promotional pricing through December 31, 2026 is $0.75 per million input tokens, $0.075 per million cached-input tokens and $3.75 per million output tokens. Starting January 1, 2027, the Standard rates are $1.50, $0.15 and $7.50. Regions and service tiers can differ.

What is the Gemini 3.8 Flash context window?

The model accepts up to 1,048,576 input tokens and can return up to 65,536 output tokens. Long context can add latency and cost, so retrieval, caching and trimming still matter.

Can Gemini 3.8 Flash inspect images and video?

Yes. It accepts text, image, video, audio and PDF input and returns text. It can analyze media but does not natively generate images.

Does Gemini 3.8 Flash work with the OpenAI SDK?

Yes. Point the OpenAI client at Google's v1beta OpenAI-compatible base URL and set the model to gemini-3.8-flash. The compatibility layer remains beta; test streaming, tool calls and error handling before launch.

Is Gemini 3.8 Flash Cyber publicly available?

Not as a normal public API model. Flash Cyber is available through the Fairwind program to approved, trusted cyber-defense organizations under additional controls.

Should every Gemini 3.7 Flash workload upgrade now?

No. Put 3.8 into tests for agents, coding and complex multimodal jobs. Keep 3.7 where it already meets requirements or uses fewer tokens, then canary the new model against task success, latency and full cost.

Sources and verification notes

Model specifications, prices, migration steps and benchmark numbers in this article come from Google sources:

• Gemini 3.8 Flash and Flash Cyber launch post
• Gemini 3.8 Flash API model page
• Latest-model and migration guide
• Gemini Enterprise Agent Platform pricing
• Gemini 3.8 Flash model card

The UI frames come from videos attached to Google's launch post and are included for commentary with attribution. Benchmarks are results reported by Google or the evaluation providers named in its table; Mengbi did not rerun them with a paid API key and received no sponsorship from Google. Pricing, model status and preview features can change, so check the official pages before a purchase or production rollout.

Last verified: September 5, 2026.