Skip to content
Gemini Link
English简体中文繁體中文Čeština한국어Türkçe

Gemini 3.8 Flash

Gemini 3.8 Flash is Google DeepMind’s high-performance multimodal Flash model for fast reasoning, coding, agentic workflows and long-context tasks.

Gemini 4 Argon, 3.8 Flash and 3.7 Flash are available
Billed by actual token usage

Examples

Input

Build a playable, retro DOS-style version of Google Maps from a single prompt.

Output

In Google's published demonstration inside the Antigravity environment, Gemini 3.8 Flash produced a working, playable DOS-style take on Google Maps from one prompt, one of several single-shot build demos (a game and a terrain visualizer built on USGS data) that Google used to illustrate its agentic coding gains.

Single-prompt build demonstration recorded in Google's announcement. This is a vendor-documented result, not a benchmark run on this site.

Google DeepMind announcement, September 2, 2026 — vendor-documented demonstration, not run on this site · Public vendor documentation, summarized for factual reference; no third-party result is presented as this site's own test.

What Gemini 3.8 Flash is

Google DeepMind released Gemini 3.8 Flash on September 2, 2026 as its most intelligent Flash model in generally available form, positioned for long-horizon software engineering, autonomous agents and complex enterprise workflows. It keeps the speed-and-cost profile of Gemini 3.7 Flash while improving on every published benchmark, and Google describes the design as "works harder": on hard tasks it runs extra reasoning passes and iterates with tools, which can use more tokens per task in exchange for a higher success rate. In this workspace it appears as a callable, priced chat mode; runtime capabilities and the server quote remain authoritative.

Specifications

  • 1,048,576 input context window (tokens)
  • 65,536 maximum output tokens
  • March 2026 knowledge cutoff
  • Sep 2, 2026 general availability

Google API introductory pricing

  • $0.75 input per 1M tokens introductory, through Dec 31, 2026
  • $3.75 output per 1M tokens introductory, through Dec 31, 2026
  • $1.50 / $7.50 from Jan 1, 2027 the introductory price doubles
  • ~90% off cached input reads $0.075 per 1M tokens on Google AI Studio

Capabilities at a glance

The Gemini API lists a broad capability set for this model. Which controls this workspace exposes at any moment is decided by the live catalog, so treat the cards below as the model's documented abilities rather than a promise about every session's options.

These figures are Google's published API prices, not this site's charges. What you pay here is always determined by the server quote before submission and the task ledger after completion.

Multimodal input

Accepts text, image, video, audio and PDF inputs; output is text.

Three thinking levels

Low, medium and high effort are supported. Minimal thinking is not supported on this model.

Tools and code

Function calling, tool use, structured outputs and server-side code execution.

Grounded answers

Search grounding and URL context let the model cite live sources.

Long-context work

The 1M-token window fits document-heavy and multi-file engineering tasks.

Agentic operation

Built for agent loops, with computer use available in preview.

Published benchmarks

Google's announcement includes a side-by-side evaluation table against Gemini 3.7 Flash and competing frontier models. Highlights: 73.7% on DeepSWE v1.1 for long-horizon software engineering (Claude Opus 5: 74.0%, GPT-5.6 Sol: 72.7%), 89.4% on Terminal-Bench 2.1 for agentic terminal coding, and clear gains over 3.7 Flash on finance and legal agent benchmarks. These are vendor-reported results published with Google's evaluation methodology, not independent audits and not tests run on this site.

BenchmarkWhat it measuresGemini 3.8 FlashGemini 3.7 FlashClaude Opus 5GPT-5.6 Sol
DeepSWE v1.1Long-horizon software engineering73.7%65.3%74.0%72.7%
Terminal-Bench 2.1Agentic terminal coding89.4%85.8%89.1%88.8%
Vals Finance Agent v2Financial analyst tasks61.4%59.0%58.6%53.8%
Harvey's Legal Agent BenchmarkComplex legal workflows10.0%8.8%6.7%2.5%
HLE-VerifiedMultidisciplinary expert reasoning54.9%53.6%54.4%54.5%
CharXiv ReasoningSynthesis from complex charts86.2%84.5%83.7%85.8%
LVBenchLong video understanding87.1%85.4%75.4%82.1%
Google's published evaluation table comparing Gemini 3.8 Flash with Gemini 3.7 Flash, Claude Opus 5, Claude Sonnet 5, GPT-5.6 Sol and GPT-5.6 Terra across pricing and thirteen benchmarks
Google's full published comparison table: pricing and benchmark results for Gemini 3.8 Flash.
DeepSWE v1.1 scatter plot of task success versus average cost per task, with Gemini 3.8 Flash at the top right among the most efficient models
DeepSWE v1.1: success rate against average cost per task. Gemini 3.8 Flash sits near the top of the efficiency frontier.

Both figures reproduce Google's published evaluation materials for Gemini 3.8 Flash; methodology notes are included in the original announcement. blog.google ↗

How to use it in this workspace

The chat workspace on this site calls Gemini 3.8 Flash as a priced mode from the live catalog. Four habits keep sessions predictable.

  1. Open the chat workspace and pick Gemini 3.8 Flash in the model selector.
  2. State the outcome you need and include only the context the task requires; long documents fit the 1M-token window, but focused input still produces cheaper, tighter answers.
  3. Check the server quote before submitting. The estimate reflects the selected runtime settings, and the task ledger records the final deduction.
  4. Review the answer before reuse, paying extra attention to numbers, dates and citations.

Prompt patterns that work well

Two patterns exploit what this model was built for: extra reasoning passes on hard tasks, and long-context synthesis. Copy them into the workspace and adapt the bracketed parts.

Long-document extraction
Read the attached document end to end. Extract every [clause / metric / deadline] that matches the rule below, list them in document order with the section they came from, and flag any conflict between sections. Rule: [your rule].
Stepwise engineering task
Implement [feature] in [repository/file]. First list the files you need to touch and the order of changes, then make each change, then summarize how to verify the result. If any step depends on an assumption, state the assumption before coding.

Because the model can spend more reasoning tokens on hard tasks, watch the quote on exploratory work: lowering the thinking level, or splitting a large task into reviewable steps, usually trims cost without hurting the outcome.

How it relates to Gemini 4 Argon

Gemini 4 Argon, Google's new frontier model, opened worldwide through the Gemini API and subscription channels on October 13, 2026. Its introductory API price is $2 / $10 per million tokens; Gemini 3.8 Flash keeps the Flash tier at $0.75 / $3.75. This workspace calls Gemini 3.8 Flash today. Argon appears here only when the runtime catalog lists it as callable and priced.

Questions

Is this a separate Gemini 4 Argon model?

No. Gemini 4 Argon is the site name; this page opens the Gemini 3.8 Flash chat mode currently available on this site.

Where is the final cost recorded?

The server quote provides the estimate and the server ledger records actual usage.

Is Gemini 3.8 Flash the same as Google's Gemini 4 Argon?

Gemini 4 Argon is officially available. This workspace calls it in this workspace when the live catalog returns gemini-4-argon-chat as available, callable and priced.

When will this site offer a Gemini 4 Argon model?

There is no schedule. A model appears here only when the site operator enables it and the runtime catalog lists it as callable and priced for this site.

What is the context window?

Google publishes a 1,048,576-token input window with up to 65,536 output tokens, and a March 2026 knowledge cutoff. The workspace may cap visible length per session; the live catalog is authoritative for what a session exposes.

What does Gemini 3.8 Flash cost?

Google's API lists an introductory $0.75 per million input tokens and $3.75 per million output tokens through December 31, 2026, doubling to $1.50 and $7.50 on January 1, 2027. Charges inside this workspace are set by the server quote and settled in the task ledger, which may differ from Google's list price.

Can it read images, PDFs, audio or video?

The Gemini API accepts text, image, video, audio and PDF inputs for this model and returns text. Which input types are enabled in this workspace depends on the live catalog; check the controls shown in the session.

What is Gemini 3.8 Flash Cyber?

A companion variant tuned for vulnerability discovery and patching, announced the same day. Google distributes it to vetted defenders through the Fairwind Program rather than the general API, so it is not the model this page serves.

Explore next