Artificial intelligenceLISTIFY team · September 26, 2026 · 12 min read

Gemini 3.8 Flash: the technical guide to benchmarks, API, pricing and how it compares

Gemini 3.8 Flash is Google's newest model, and on coding and agentic tasks it reaches the level of large models at a fraction of their price. It launched on September 2, 2026 as the third Flash model in six weeks. It scores 73.7% on DeepSWE, practically the same as Claude Opus 5, while an input token costs about a seventh as much.

Gemini 3.8 Flash article cover: a 1 million token context window, 73.7% on DeepSWE v1.1 and $0.75 and $3.75 per million tokens

The newest Pro model, meanwhile, is still Gemini 3.1 Pro from February 19, 2026. This guide covers the specs, benchmarks, pricing including the January 2027 increase, the Cyber variant and how it compares with the competition.

01What is Gemini 3.8 Flash

Gemini 3.8 Flash is a fast, low-cost model from Google DeepMind, which Google calls its smartest Flash model yet. It is built for scaling AI agents: coding, agentic tasks and multi-step reasoning. It shipped alongside Gemini 3.8 Flash Cyber, a variant for cyber defense. (Google AI for Developers)

“Gemini 3.8” is just the generation label. The generally available model is gemini-3.8-flash; there is no separate Gemini 3.8 Pro model yet. (Agentpedia)

Google is shipping Flash models in quick succession: 3.6 Flash, 3.7 Flash and 3.8 Flash within six weeks. Each version builds on the agentic capabilities of the previous one. (Agentpedia)

Independent evaluation

On the current Intelligence Index v4.3, Gemini 3.8 Flash scores 41 at high reasoning, 2 points more than 3.7 Flash (39); GPT-5.6 Sol at xhigh scores 44. The model outputs roughly 330 tokens per second. (Artificial Analysis, as of September 26, 2026) According to Artificial Analysis' launch article, it sits on the Pareto frontier of intelligence and cost per task. (Artificial Analysis)

02The Gemini model family in September 2026

Gemini 3.8 Flash is the newest model, but the strongest Pro version is the older Gemini 3.1 Pro, which is still only a preview in the API (gemini-3.1-pro-preview). Current models:

Model	Released	Focus
Gemini 3.8 Flash	September 2, 2026	Fast agentic model, coding, documents, video
Gemini 3.8 Flash Cyber	September 2, 2026	Finding and fixing vulnerabilities, only through the Fairwind program
Gemini 3.5 Flash-Lite	July 21, 2026	Cheapest high-volume tasks
Gemini 3.1 Pro	February 19, 2026	Strongest Pro version, a preview in the API
Gemini 3 Deep Think	February 12, 2026	Deep reasoning for hard problems

Sources: Wikipedia: Google Gemini, Capital and Compute, Google AI for Developers

In September, Google also released the Gemini app for Windows and introduced agentic video understanding. (The Soo Group)

03Gemini 3.8 Flash technical specifications

Gemini 3.8 Flash has a 1 million token context window, up to 65,536 output tokens and an explicit reasoning mode.

Spec	Value
Developer	Google DeepMind
API model ID	`gemini-3.8-flash`
Release	September 2, 2026 (stable, generally available)
Predecessor	Gemini 3.7 Flash
Context window	1M tokens
Max output	65,536 tokens
Reasoning	Explicit mode; adjustable level and thinking budget
Multimodality	Text, images, audio, video and documents (including agentic processing of long video)
Output speed	About 300 tokens/s (high reasoning, measured at launch)
Average time per task	About 2.5 min (high reasoning)
Price (introductory, through December 31, 2026)	$0.75 input / $3.75 output per 1M tokens
Price (from January 1, 2027)	$1.50 input / $7.50 output per 1M tokens
Availability	Gemini API, Google AI Studio, Gemini app (Pro and Ultra subscriptions)

Sources: Agentpedia, BenchLM, Artificial Analysis, Google DeepMind: model card

Reasoning levels

On index v4.3, the model scores 41 at high reasoning, 40 at medium and 33 at low. An index task costs about $1.24 at high and about $0.93 at medium. (Artificial Analysis)

Agentic video understanding

Instead of fixed sampling at 1 frame per second, the model decides which parts of a video to look at. On the LVBench long-video benchmark, that yields 87.8% versus 87.1% without agentic mode. (Agentpedia)

04How to use Gemini 3.8 Flash

You can run Gemini 3.8 Flash in the Gemini app, in Google AI Studio or through the Gemini API.

In Google AI Studio

Open Google AI Studio and sign in with your Google account.
Select Gemini 3.8 Flash in the model settings.
Set the reasoning level and tools (search, code execution) to fit the task.
You can export a finished prompt as API code.

Through the Gemini API

An illustrative call in Python with the google-genai SDK:

python

from google import genai

client = genai.Client()  # reads GEMINI_API_KEY from the environment

response = client.models.generate_content(
    model="gemini-3.8-flash",
    contents="Find and fix the failing test in this repository.",
)

print(response.text)

Example source: Agentpedia. Reasoning level, thinking budget and tools are set the same way as for 3.7 Flash; check the exact parameter names in the Gemini API documentation.

Migrating from 3.7 Flash: just change the model name. Agentic video settings stay the same. (Agentpedia)

05Gemini 3.8 Flash benchmarks

Gemini 3.8 Flash wins agentic and domain tests (code, terminal, finance, law, biology), but trails the large models on computer use and general knowledge work. Numbers from Google's official card:

Benchmark	What it measures	Gemini 3.8 Flash	Gemini 3.7 Flash	Claude Opus 5	GPT-5.6 Sol
DeepSWE v1.1	Long software tasks	73.7%	65.3%	74.0%	72.7%
Terminal-Bench 2.1	Terminal work	89.4%	85.8%	89.1%	88.8%
Terminal-Bench 4.0	General agent	19.1%	11.2%	51.8%	37.3%
Vals Finance Agent v2	Finance agent	61.4%	59.0%	58.6%	53.8%
Harvey Legal Agent	Legal agent	10.0%	8.8%	6.7%	2.5%
GDPval-AA v2 (Elo)	Knowledge work	1545	1482	1824	1710
HLE-Verified	Expert reasoning	54.9%	53.6%	54.4%	54.5%
CharXiv Reasoning	Reading charts	86.2%	84.5%	83.7%	85.8%
LVBench	Long video	87.8%	85.4%	75.4%	82.1%
OSWorld 2.0	Computer use	59.0%	50.6%	75.4%	62.6%
GDP.PDF	Document understanding	35.0%	34.0%	37.0%	40.0%
LABBench2	Biology research	86.2%	82.1%	84.2%	82.1%

Source: Google's official card via Agentpedia. All figures come from the vendor, and Google chose the competitors' configurations. Claude Opus 5.5 replaced Claude Opus 5 on September 22, 2026 and is not yet on Google's card.

DeepSWE v1.1: long software tasks

Claude Opus 574.0%
Gemini 3.8 Flash73.7%
GPT-5.6 Sol72.7%
Gemini 3.7 Flash65.3%

Google's official card; Google chose the competitors' configurations. Source: Agentpedia

The strongest argument is efficiency: on DeepSWE, Gemini 3.8 Flash matches Opus 5's accuracy at a fraction of the cost per task. Glean reports that in long document workflows it completed more than three times as many tasks as 3.7 Flash. (Google DeepMind)

OSWorld 2.0: computer use

Claude Opus 575.4%
GPT-5.6 Sol62.6%
Gemini 3.8 Flash59.0%
Gemini 3.7 Flash50.6%

A weaker spot for 3.8 Flash. Source: Google's official card via Agentpedia

How it compares with the September 2026 flagships

Against GPT-6 Astra and Claude Fable 5.1, Gemini 3.8 Flash is a class below on general intelligence and agentic tasks, but it keeps pace on long software tasks and costs far less.

Benchmark	Gemini 3.8 Flash	GPT-6 Astra	Claude Fable 5.1
AA Intelligence Index v4.3	41	53	53
DeepSWE v1.1	73.7%	74.1%	67.4%
Terminal-Bench 4.0	19.1%	57.9%	55.8%
API price input / output (per 1M tokens)	$0.75 / $3.75	$10 / $50	$10 / $50

Gemini 3.8 Flash: Google's card via Agentpedia. Index: Artificial Analysis as of September 26, 2026. GPT-6 Astra and Claude Fable 5.1: OpenAI's comparison table (see our GPT-6 Astra guide), Fable 5.1 on Terminal-Bench 4.0 per Anthropic. The numbers come from different measurements, so treat them as a rough guide.

On the independent BenchLM leaderboard, 3.8 Flash scores 69.2% on CursorBench 3.2 (Fable 5.1: 73.4%) and 59% on OSWorld 2.0 (GPT-6 Astra: 72.6%). (BenchLM) For more on the competitors, see our GPT-6 Astra guide, Claude Fable 5.1 guide, Grok 4.7 guide and Muse Spark 1.3 deep dive.

For all seven September models side by side, with pricing and benchmark charts, see our big comparison of September 2026 AI models.

06Gemini 3.8 Flash Cyber and the Fairwind program

Gemini 3.8 Flash Cyber is a specialized variant for finding and automatically fixing vulnerabilities. It is not in the public API; only trusted defenders get it through the new Fairwind program. Governments and national cyber agencies, critical infrastructure operators and key technology platforms with a proven ethical track record can apply through a form. Google has not published pricing. (Google DeepMind: Fairwind)

Benchmark / deployment	Gemini 3.8 Flash Cyber result	Comparison
CyberGym Pass@1 (finding vulnerabilities in C/C++)	86.2%	GPT-5.5-Cyber 85.6%, Mythos 5 83.8%, 3.5 Flash Cyber 77.5%
CWE-Bench (fixes, pass@1)	47.2%	A leading frontier model 47.8%, at a much higher price
Chrome Security team	2.6x more correct fixes	Than the best larger commercial models
Wiz (internal penetration test)	7.5% to 9.7% higher detection	At 2.3x to 5.2x lower cost

Source: Agentpedia, following Google DeepMind's methodology.

By available reports, the Cyber variant found a bug in Chrome that engineers had missed for 13 years. (Miraflow)

Google did not test the standard Gemini 3.8 Flash on CyberGym or CWE-Bench, so do not use it as a replacement for a specialized security model. (Miraflow)

07Gemini 3.8 Flash pricing and availability

Through the end of the year, Gemini 3.8 Flash costs $0.75 per million input tokens and $3.75 per million output tokens. From January 1, 2027, the price doubles.

Model	Input (per 1M tokens)	Output (per 1M tokens)
Gemini 3.8 Flash, introductory price (through December 31, 2026)	$0.75	$3.75
Gemini 3.8 Flash from January 1, 2027	$1.50	$7.50
Claude Sonnet 5	$2.00	$10.00
GPT-5.6 Sol	$4.00	$20.00
Claude Opus 5.5	$4.00	$20.00
Claude Opus 5 (previous generation)	$5.00	$25.00
GPT-6 Astra / Claude Fable 5.1	$10.00	$50.00

Sources: Agentpedia, OpenAI, Anthropic, Claude: API pricing

Gemini 3.8 Flash price over time: introductory price of $0.75 per million input and $3.75 per million output tokens from launch on September 2, 2026 through December 31, 2026, then double at $1.50 and $7.50 from January 1, 2027 — Plan budgets around the price from January 2027. Source: Agentpedia

Plan budgets around the price from January 2027, not the introductory one. Even after the increase, 3.8 Flash is 1.3x to 6.7x cheaper on input than the competitors in the table. The same price schedule also applies to Gemini 3.7 Flash. (Agentpedia)

Input price per 1M tokens

GPT-6 Astra$10.00
GPT-5.6 Sol$4.00
Claude Opus 5.5$4.00
Claude Sonnet 5$2.00
Gemini 3.8 Flash from January 1, 2027$1.50
Gemini 3.8 Flash through 2026$0.75

Sources: Agentpedia, OpenAI, Claude: API pricing

The model is available in the Gemini API and Google AI Studio. In the Gemini app, now also on Windows, Pro and Ultra subscribers get it. (BenchLM, The Soo Group)

08Who should use Gemini 3.8 Flash

Gemini 3.8 Flash pays off wherever you run a lot of agentic tasks and cost per task matters. For the hardest general tasks, reach for a larger model.

Situation	Recommendation	Why
Coding agents, test fixes, CI	Gemini 3.8 Flash	DeepSWE on par with Opus 5 at a fraction of the price
Finance, law, biology	Gemini 3.8 Flash	Leads Vals Finance, Harvey Legal and LABBench2
Long documents and videos	Gemini 3.8 Flash	1M tokens, agentic video processing
High volumes of simple queries	Gemini 3.5 Flash-Lite	The cheapest option
Computer use, general agent	GPT-6 Astra or Claude Opus 5.5	3.8 Flash scores only 59% on OSWorld and 19% on Terminal-Bench 4.0
Hardest code and research	Claude Fable 5.1 or GPT-6 Astra	Higher general intelligence
Deep reasoning in the Google ecosystem	Gemini 3.1 Pro or 3 Deep Think	Google's strongest models
Cyber defense	Gemini 3.8 Flash Cyber (Fairwind)	Specialized model

The recommendations are based on the benchmarks above (Agentpedia). Always test the model on your own tasks before migrating production.

09How to prompt Gemini 3.8 Flash

Gemini 3.8 Flash is built for agentic loops, so give it the goal and a way to verify the result, not individual steps.

Start at medium reasoning. On index v4.3, medium scores 40 instead of 41, but a task costs about a quarter less ($0.93 vs $1.24); switch to high only where the result falls short. Leave low (33 points) for simple tasks. (Artificial Analysis)
Define what done looks like. For example: “all tests pass, no new dependencies.” The model then iterates until it meets the condition.
Use the long context. Load the whole document, contract or video instead of excerpts; with 1M tokens the model can handle all of it.
Turn on agentic processing for video. The model picks the important segments itself instead of sampling at a fixed rate.
Watch the output length. The limit is 65,536 tokens, so split long generation (whole modules, reports) into several steps.
Measure cost per task, not per token. A higher reasoning level makes tasks longer; the average at high is about 2.5 minutes.

10Gemini 3.8 Flash FAQ

What is Gemini 3.8 Flash?

A fast, low-cost model from Google DeepMind released on September 2, 2026, focused on coding, agentic tasks and multi-step reasoning. Google calls it its smartest Flash model yet.

How much does Gemini 3.8 Flash cost?

Through December 31, 2026, the introductory price is $0.75 per million input tokens and $3.75 per million output tokens. From January 1, 2027, it will be $1.50 and $7.50.

How big is the Gemini 3.8 Flash context window?

1 million input tokens and up to 65,536 output tokens.

Is there a Gemini 3.8 Pro?

Not yet. The newest Pro version is Gemini 3.1 Pro from February 19, 2026, still a preview in the API.

Is Gemini 3.8 Flash better than Claude or GPT?

On long software tasks (DeepSWE) it reaches the level of Claude Opus 5 and GPT-6 Astra at a fraction of the price. On computer use and general intelligence it trails them.

What is Gemini 3.8 Flash Cyber?

A variant for finding and fixing vulnerabilities, available only to trusted defenders through the Fairwind program. It scores 86.2% on CyberGym.

11Verdict: is Gemini 3.8 Flash worth it?

In September 2026, Gemini 3.8 Flash offers the best price to performance for agentic coding and work with long documents. On code it reaches the large models at a fraction of their price.

It is not a universal winner, though. GPT-6 Astra and the Claude models have the edge on computer use and general agentic tasks, and the introductory price ends in January 2027. If you are scaling agents and watching costs, start with 3.8 Flash; keep a larger model in reserve for the hardest tasks.

Want to put AI agents to work in your own processes? We help you choose the model and connect it to your systems as part of our AI implementation service.

12Sources

LISTIFY teamWebsites, apps and marketing from Prague since 2008

All articles →

Artificial intelligenceSeptember 27, 2026 · 13 min read