Gemini 3.8 Flash

gemini-3.8-flash

Google's latest Flash-level tool, suitable for coding, agent workflows, knowledge work, and multimodal reasoning, reportedly with higher accuracy than Gemini 3.7 Flash and a context window of 1 million tokens.

Context window1.0M

ProviderGemini

Released2026/09/04

Playground

Pricing

The same model is available through multiple service channels — choose based on latency, reliability and cost.

Prices in $ / 1M tokens

To pick a channel, add a provider field to the request body, for example "provider": { "channel": "direct" }. Valid values are direct / stable / economical; omit it to use the default channel.

Direct

Direct upstream connection — best when you need native behavior and the full context window.

Input	Output	Cache read
0.75/M	3.75/M	0.07/M

Overview

Input

Text Image

Output

Text

Gemini 3.8 Flash API: A Flash-Tier Model That Puts Professional Agent Tasks Above Flagships

Gemini 3.8 Flash is a multimodal model released by Google on September 2, 2026, just three weeks after 3.7 Flash—making it the third Flash model in six weeks. It continues the cost positioning of the Flash tier, but its gains in coding and agent capabilities are rare for this series: DeepSWE v1.1 rose from 65.3% to 73.7%, and OSWorld 2.0 from 50.6% to 59.0%, both up by 8.4 percentage points.

Its design philosophy is "willing to spend more effort": when facing complex tasks, it performs extra reasoning steps and repeatedly calls tools before giving an answer. The cost is higher token usage at higher levels; the payoff is that on vertical agent benchmarks such as finance and law, it directly outperforms much larger flagship models—scoring 61.4% on Vals Finance Agent v2 (Claude Opus 5 58.6%, GPT-5.6 Sol 53.8%) and 10.0% on the Harvey legal agent (Opus 5 6.7%, Sol 2.5%).

Note that 3.8 Flash introduces backward-incompatible API changes: temperature, top_p, top_k, and candidate_count are removed, and thinking_budget is replaced by the string enum thinking_level. Code migrating from older versions needs to be updated.

SeaWhale AI provides Gemini 3.8 Flash through an OpenAI-compatible interface, supporting tool calling, thinking-level control, streaming output, and multimodal input.

Get API Key · Model ID: gemini-3.8-flash

Why Choose Gemini 3.8 Flash

Beats flagships on vertical agents — Vals Finance Agent v2 61.4%, Harvey legal 10.0%, both higher than Opus 5 and GPT-5.6 Sol
Major coding leap — DeepSWE v1.1 up from 65.3% on 3.7 Flash to 73.7%
Computer use up 8.4 points — OSWorld 2.0 from 50.6% to 59.0%
Terminal-Bench 2.1 89.4% — slightly above Opus 5's 89.1% and Sol's 88.8%
Leads in long-video understanding — LVBench 87.8%, 12 points higher than Opus 5's 75.4%
1M token context + full-modality input — text, image, audio, video, PDF; max output 64K tokens

Core Capabilities

01 Long-Horizon Coding

DeepSWE v1.1 73.7%, 8.4 points higher than 3.7 Flash, and only 0.3 points behind Claude Opus 5's 74.0%—the first time a Flash-tier model has kept pace with a flagship on long-horizon software engineering. Terminal-Bench 2.1 89.4%, also slightly ahead of Opus 5 and GPT-5.6 Sol.

End-to-end solving of complex engineering problems
Multi-step execution in real terminal environments
Codebase-level refactoring and defect repair

02 Vertical-Domain Agents

This is the most counterintuitive part of 3.8 Flash: on professional agent benchmarks in finance and law, it directly surpasses flagship models in much higher price tiers. Vals Finance Agent v2 61.4% (Opus 5 58.6%, Sol 53.8%), Harvey legal agent 10.0% (Opus 5 6.7%, Sol 2.5%). GDPval-AA v2 also rose from 1482 Elo to 1545.

Financial data analysis and reporting agents
Legal document processing and compliance workflows
Multi-step orchestration of real professional tasks

03 Full-Modality Understanding

Supports text, image, audio, video, and PDF input, with text output. LVBench 87.8% (3.7 Flash 85.4%, Opus 5 only 75.4%); long-video understanding is the multimodal area where it opens the widest gap. CharXiv chart reasoning 86.2%.

Long-video content understanding and retrieval
Scientific chart and data-visualization interpretation
Unified processing of mixed materials such as PDF and audio

04 Thinking Levels and Tool Orchestration

thinking_level supports three tiers: low, medium, and high, with medium as the default (this model does not support minimal). At higher tiers, the model proactively increases reasoning steps and repeatedly calls tools. Built-in tools cover Google Search, Google Maps, code execution, URL context, file retrieval, and computer use.

Switch thinking depth based on task difficulty
Iterative tool calls and result verification
Structured output and function calling

Best-Use Scenarios

Scenario	Description
Financial analysis agents	Vals Finance Agent v2 61.4%, above Opus 5 and GPT-5.6 Sol
Legal document agents	Harvey legal benchmark 10.0%, highest among compared models
Long-horizon coding	DeepSWE v1.1 73.7%, only 0.3 points behind Opus 5
Long-video understanding	LVBench 87.8%, 12 points higher than Opus 5
Biomedical research	LABBench2 86.2%, BioMysteryBench hard tier 43.5%→56.5%
High-concurrency production traffic	Run professional agent tasks at Flash-tier cost

Differences Between Gemini 3.8 Flash, Gemini 3.7 Flash, and Claude Opus 5

Capability	Gemini 3.8 Flash	Gemini 3.7 Flash	Claude Opus 5
Model ID	`gemini-3.8-flash`	`gemini-3.7-flash`	`claude-opus-5`
Vendor	Google	Google	Anthropic
Release date	2026-09-02	2026-08-13	—
DeepSWE v1.1	73.7%	65.3%	74.0%
Terminal-Bench 2.1	89.4%	85.8%	89.1%
Terminal-Bench 4.0	19.1%	11.2%	51.8%
OSWorld 2.0	59.0%	50.6%	75.4%
Vals Finance Agent v2	61.4%	59.0%	58.6%
Harvey legal agent	10.0%	8.8%	6.7%
HLE-Verified	54.9%	53.6%	—
LVBench	87.8%	85.4%	75.4%
Context window	1M tokens	1M tokens	—
Max output	64K tokens	—	—
Input modalities	Text/Image/Audio/Video/PDF	Same as left	Text/Image
Thinking levels	`low`/`medium`/`high`	`thinking_budget` numeric	Adjustable
Positioning	Primary Flash-tier model for professional agents	Previous-generation Flash mainstay	General-purpose flagship

For specific billing, refer to the real-time price card at the top of the page.

FAQ

When was Gemini 3.8 Flash released? Released on September 2, 2026, only three weeks after 3.7 Flash (August 13). It is the third Flash model in six weeks. A restricted-access Gemini 3.8 Flash Cyber was also released at the same time.

What has been upgraded compared with Gemini 3.7 Flash? All benchmarks reported by Google rose; the largest gains were in coding and computer use: DeepSWE v1.1 +8.4 points (65.3%→73.7%), OSWorld 2.0 +8.4 points (50.6%→59.0%), Terminal-Bench 4.0 +7.9 points. Biomedical hard-tier gains were also large, with BioMysteryBench's human-hard set going from 43.5% to 56.5%. Knowledge gains were much smaller: HLE-Verified was only +1.3 points.

Are there areas where it lags behind? Yes, and it is structural. It trails noticeably on the hardest general agent tests: Terminal-Bench 4.0 is only 19.1%, while Opus 5 is 51.8% and GPT-5.6 Sol is 37.3%; its 59.0% on OSWorld 2.0 also falls below Opus 5's 75.4%. The pattern is clear—it wins on cheaper, more vertical agent tasks, and loses on open-ended high-difficulty long-horizon tasks. In addition, knowledge gains are limited: GDP.PDF 35.0% is below Sol's 40.0%.

What are the context and output limits? Context window of 1M tokens and maximum output of 64K tokens. Input supports text, image, audio, video, and PDF; output is plain text.

How do I set the thinking level? thinking_level has three tiers: low, medium, and high, with medium as the default. This model does not support the minimal tier. Use low for tasks such as transcription retrieval and basic metadata extraction; use high for dense reasoning. Note that at the high tier the model proactively increases reasoning steps and tool calls, so token usage will rise noticeably.

Do I need to change code when migrating from older versions? Yes. 3.8 Flash removes four parameters: temperature, top_p, top_k, and candidate_count; thinking_budget is replaced by the string enum thinking_level. On the function-calling side, FunctionResponse must include the call_id and name fields, and multimodal resources must be placed in the response payload. Multi-turn sessions rely on the server-side previous_interaction_id, and prefilled model turns are no longer supported.

What is Gemini 3.8 Flash Cyber? Can I use it directly? No. It is a security-focused variant released at the same time, available only to vetted government agencies, critical-infrastructure operators, and software maintainers through the Fairwind program, with looser safety guardrails. Google says it produces 2.6 times more correct Chrome vulnerability patches than the strongest commercial model, with a success rate above 70% on real-world vulnerability discovery (covering 20 programming languages); its CWE-Bench patch pass@1 is 47.2%, roughly on par with the leading flagship's 47.8%. The standard gemini-3.8-flash does not include these relaxed capabilities.

Why Use Gemini 3.8 Flash API via SeaWhale AI

No overseas account required — direct connection from China, no need to build your own proxy
OpenAI-compatible interface — connect by changing two lines in existing code
Cross-vendor comparison — directly compare pricing and performance with Claude, GPT, and Grok under the same account
Unified billing — centrally manage usage and spending across multiple models

API

API integration

Model IDUse this value as the model in inference requests

gemini-3.8-flash

API KeyBearer token used to authenticate inference requests

Manage API keys

Base URLOpenAI compatible · /chat/completions

https://api.haijingai.com/v2/

provider OptionalSelects a service channel; omit it and the system picks the default

"provider": { "channel": "direct" }

gemini-3.8-flash usage examples

SeaWhale AI is compatible with the OpenAI API protocol, so you can call it with the OpenAI SDK or plain HTTP requests. Streaming is enabled by default.

cURLPythonNode.js

curl https://api.haijingai.com/v2/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer <API_KEY>" \
  -d '{
    "model": "gemini-3.8-flash",
    "messages": [
      {"role": "system", "content": "You are a helpful assistant."},
      {"role": "user", "content": "Hello!"}
    ],
    "provider": { "channel": "direct" },
    "stream": true
  }'
# provider is optional — remove this line to use the default channel

from openai import OpenAI

client = OpenAI(
    base_url="https://api.haijingai.com/v2",
    api_key="<API_KEY>",
)

stream = client.chat.completions.create(
    model="gemini-3.8-flash",
    messages=[
        {"role": "system", "content": "You are a helpful assistant."},
        {"role": "user", "content": "Hello!"},
    ],
    stream=True,
    # Optional: pick a service channel; omit to use the default
    extra_body={"provider": {"channel": "direct"}},
)

for chunk in stream:
    if chunk.choices[0].delta.content:
        print(chunk.choices[0].delta.content, end="", flush=True)

import OpenAI from 'openai'

const client = new OpenAI({
  baseURL: 'https://api.haijingai.com/v2',
  apiKey: '<API_KEY>',
})

const stream = await client.chat.completions.create({
  model: 'gemini-3.8-flash',
  messages: [
    { role: 'system', content: 'You are a helpful assistant.' },
    { role: 'user', content: 'Hello!' },
  ],
  stream: true,
  // Optional: pick a service channel; omit to use the default
  // @ts-expect-error provider is a SeaWhale AI extension, not in the OpenAI SDK types
  provider: { channel: 'direct' },
})

for await (const chunk of stream) {
  process.stdout.write(chunk.choices[0]?.delta?.content ?? '')
}