HomeAI models
Google: Gemini 3.8 Flash
Google
•
Gemini 3.8 Flash (High)
•
Released 
September 2026

Google
Gemini 3.8 Flash (High)

Google's most intelligent Flash model for long-horizon coding agents and multi-step enterprise workflows at Flash-tier cost.

Modality:
Text
Image
Audio
Video
PDF
model ID
google/gemini-3.8-flash
Start for free
arrow black icon
View API Docs
arrow white icon

Output Speed *

125.73
tok/s

Intelligence Index *

40.9
/ 100

Context Window *

1048576
tokens

Input price

4.5
Anytoken

Output price

22.5
Anytoken
Gemini 3.8 Flash: Near-Frontier Agentic Coding at Flash-Tier Cost Gemini 3.8 Flash is Google's latest Flash-tier model, built on Gemini 3.7 Flash within the Gemini 3 family. It sits between the deep-reasoning Pro models and the high-throughput Flash-Lite line, positioned as the primary agentic workhorse. Google describes it as its most intelligent Flash model, engineered for long-horizon software engineering, autonomous agents, and complex enterprise workflows. Its key characteristic is a design that "works harder"—executing more reasoning steps and iterative tool calls to approach frontier-model coding accuracy while retaining Flash speed and pricing. It fits agentic coding and document-heavy knowledge work best. Start building with the Gemini 3.8 Flash API on AnyAPI.ai

Performance

Where Gemini 3.8 Flash's "Work Harder" Design Pays Off

Gemini 3.8 Flash is strongest in agentic coding and long-horizon software engineering. Google reports 89.4% on Terminal-Bench 2.1 (up from 85.8% for 3.7 Flash) and 73.7% on DeepSWE v1.1 (up from 65.3%), placing it within a point of frontier models like Claude Opus 5 on those tasks. These gains come from a deliberate design choice: the model executes more reasoning steps and iterative tool calls. The consequence is meaningful—agentic coding accuracy approaches near-frontier levels at Flash-tier speed and cost, but per-task token consumption and turn counts rise on complex work.

Benchmarks

Independent Signals: Coding Gains Are Real, Reasoning Is Flatter

Independent testing supports the coding improvement but shows an uneven picture. On the public DeepSWE v1.1 leaderboard run with a common agent harness, Artificial Analysis and DataCurve report roughly 74% at high effort versus about 65% for 3.7 Flash. That same run shows the cost of the gain: the 3.8 Flash run averaged around 143,000 output tokens and 166 agent steps, versus 107,000 tokens and 125 steps for 3.7 Flash. Exam-style reasoning (HLE) barely moved, and harder agent benchmarks like Terminal-Bench 4.0 and OSWorld-2.0 remain well behind top frontier models. Keep your own eval set.

Output Speed

*
125.73
tok/s

Intelligence Index

*
40.9
/ 100

MMLU *

Broad world knowledge and problem-solving
0
%

GPQA *

PhD-level scientific reasoning across physics, biology, chemistry.
95
%

HLE *

Adherence to multi-step structured instructions.
48
%

LiveCodeBench *

Tool-calling reliability in long agentic loops.
0
%

Technical Specifications

What the model supports

Gemini 3.8 Flash accepts text, image, audio, and video input and returns text only, with a 1,048,576-token context window and a 65,536-token maximum output. Reasoning is controlled through three thinking levels—LOW, MEDIUM (default), and HIGH—with MINIMAL explicitly unavailable and returning a validation error. The high output ceiling and 1M context suit multi-file code and document-heavy agent runs, but the mandatory thinking floor means every request carries some reasoning overhead, which increases token usage on complex tasks. It is generally available for production under the model ID gemini-3.8-flash.
Verified Specifications — 
Gemini 3.8 Flash (High)
*
Input modalities
Text
Image
Audio
Video
PDF
output modalities
Text
Context window
1048576
 tokens
Maximum output tokens
65536
Reasoning
Yes
Knowledge cutoff
September 2026
Pricing (standard)
4.5
 AnyTokens in
 / 
22.5
 AnyTokens out

Limitations & Trade-offs

Where Gemini 3.8 Flash (High) falls short

1
Higher per-task cost from token consumption. Because 3.8 Flash "works harder," it executes more reasoning steps and iterative tool calls, and reports about 40% higher cost per task than 3.7 Flash despite identical per-token pricing. Independent DeepSWE runs show roughly 143,000 output tokens and 166 agent steps versus 107,000 and 125 for 3.7 Flash. This matters for cost-sensitive, high-volume workloads where efficiency dominates. For those, use lower thinking levels or 3.7 Flash instead.
2
No MINIMAL thinking level. Unlike some siblings, 3.8 Flash does not support thinking_level="MINIMAL"; setting it returns an API validation error, and the supported values are LOW, MEDIUM (default), and HIGH. Every request therefore carries at least some reasoning overhead. Latency- and cost-critical, high-throughput use cases that previously relied on a minimal thinking floor cannot replicate that behavior here and may be better served by a lighter model.
3
Uneven gains beyond coding. Improvements cluster in coding, terminal, agentic, and multimodal understanding tasks, while exam-style reasoning barely moved—HLE-Verified edged up to 54.9%. Harder agent and computer-use benchmarks remain well behind top frontier models, with Terminal-Bench 4.0 near 19.1% and OSWorld-2.0 at 59.0% versus much higher Opus 5 figures. For hard, open-ended reasoning or the most demanding general agents, a frontier Pro-tier model is preferable.
4
Text-only output and knowledge-cutoff nuance. Despite accepting text, image, audio, and video input, 3.8 Flash returns text only—no native image, audio, or video generation. Its knowledge cutoff is also uneven: March 2026 for some domains while others may be limited to January 2025. Applications needing generated media output or uniformly recent knowledge should pair it with dedicated generation models or grounding tools.

Best-Fit Workloads

Where this model earns its place

01

Long-horizon coding agents

‍
Gemini 3.8 Flash is engineered for autonomous, multi-step software engineering. Google reports 89.4% on Terminal-Bench 2.1 and 73.7% on DeepSWE v1.1, with independent leaderboards confirming a clear gain over 3.7 Flash. Its 1M-token context holds large multi-file repositories and its 64K output supports substantial code generation. This fits agentic IDE assistants and CI-integrated issue-resolution bots—though budget for higher token and turn counts on complex tasks.

02

Professional knowledge-work agents

‍
The model shows measured gains on domain-specific professional benchmarks, outperforming 3.7 Flash and several frontier models on Vals Finance Agent v2 and Harvey's Legal Agent Benchmark. Combined with strong long-document handling, it suits finance, legal, and analytical agents that read large filings, reason across multiple steps, and call tools iteratively—delivering near-frontier accuracy on these verticals at Flash-tier cost.

03

Long-context document and multimodal processing

‍
With a 1,048,576-token context window and text, image, audio, and video input, Gemini 3.8 Flash processes large documents and mixed-media inputs in a single request. Google reports strong results on chart reasoning (CharXiv) and long-video understanding (LVBench). This fits RAG over large corpora, video and chart analysis, and document-heavy workflows, where Google reports it completing over three times as many long-running tasks as 3.7 Flash.

04

High-volume production agents needing cost efficiency

‍
As Google's agentic Flash workhorse, 3.8 Flash targets cost-effective scaling of general-purpose, production-ready agents. Its Flash-tier per-token pricing makes it attractive as a default for high-volume deployments, but its "work harder" design raises per-task cost—so tune thinking levels down and monitor token usage, or fall back to 3.7 Flash where efficiency is the hard constraint.

Pricing in anytokens via AnyAPI
Input
4.5
₳
Output
22.5
₳
Cache write
0.25
₳
Cache read
0.45
₳
Get API access
arrow black icon
Read Docs

on this page

PerformanceBenchmarksTechnical SpecificationsLimitations & Trade-offsBest-Fit WorkloadsIntegrationFAQ

Integration

Access Gemini 3.8 Flash (High) via AnyAPI.ai

Access Gemini 3.8 Flash (High) through AnyAPI.ai using a unified API built for multi-model AI applications. Integrate Gemini 3.8 Flash (High) without maintaining a separate provider-specific connection, and keep the flexibility to test, switch, or combine models as your application requirements evolve.

01

One API integration

Access Gemini 3.8 Flash (High) and other AI models through the same API workflow instead of maintaining separate integrations for every provider.

02

Easy model switching

Test Gemini 3.8 Flash (High) against alternative models or switch models as your performance, capability, or cost requirements change without rebuilding your application around another provider API.

03

Flexible for production

Use Gemini 3.8 Flash (High) from experimentation through production while keeping your AI stack flexible as workloads, traffic, and model requirements evolve.

04

Multi-model applications

Use Gemini 3.8 Flash (High) for the workloads where it performs best and combine it with other models for tasks that require different capabilities, performance, or efficiency.

Start free
arrow black icon
Read Docs

Frequently Asked Questions

Answers to common questions about integrating and using this AI model via AnyAPI.ai

Gemini 3.8 Flash is Google's most intelligent Flash-tier model, tuned for long-horizon software engineering, autonomous agents, and complex enterprise knowledge work. It reports strong agentic coding results, including 89.4% on Terminal-Bench 2.1 and 73.7% on DeepSWE v1.1, approaching frontier-model accuracy at Flash-tier speed and cost.

Gemini 3.8 Flash supports a 1,048,576-token (1M) context window and a maximum output of 65,536 tokens. It accepts text, image, audio, and video input but returns text only. These limits are identical to Gemini 3.7 Flash.

Gemini 3.8 Flash is built on 3.7 Flash with the same specs and current introductory pricing, but is tuned to "work harder," delivering higher coding and agentic accuracy at the cost of more tokens and agent turns per task—Google cites roughly 40% higher cost per task. It also removes the MINIMAL thinking level.

Yes. Gemini 3.8 Flash exposes tunable thinking levels—LOW, MEDIUM (default), and HIGH—to trade latency and token cost against depth. Unlike some Flash models, it does not support MINIMAL; setting that value returns an API validation error.

Gemini 3.8 Flash has a knowledge cutoff of March 2026 for some domains, while in other domains its knowledge may be limited to January 2025. For uniformly recent information, pair it with grounding or web-search tools.

* Benchmark data source: Artificial Analysis artificialanalysis.ai