Gemini 3.8 Flash
Gemini 3.8 Flash is Google's most intelligent Flash-tier model, engineered for long-horizon software engineering, autonomous agents, and complex enterprise workflows. Featuring a 1M-token context window and up to 64k output tokens, it introduces tunable reasoning levels (Low, Medium, High) to balance output latency and execution depth. Built for resilient multi-step planning, deterministic tool orchestration, and complex multi-file code refactoring, Gemini 3.8 Flash minimizes iterative execution errors—powering autonomous AI agents, enterprise data pipelines, technical software refactoring, and high-concurrency API deployment.
Related Models
Gemini 3.7 Flash
Gemini 3.7 Flash is Google's flagship high-efficiency workhorse model engineered for autonomous agentic workflows and native multimodal reasoning. Built on enhanced core algorithmic foundations, it delivers ultra-low latency inference alongside robust multi-step tool execution, adaptive task planning, and end-to-end software engineering. Operating at $0.75 per million input tokens, Gemini 3.7 Flash powers enterprise agent orchestration, low-latency conversational interactive applications, real-time code generation, and high-volume data processing pipelines.
Gemini Omni Flash Text-to-Video
Gemini Omni Flash Text-to-Video is Google DeepMind's flagship unified multimodal model under the Gemini Omni series. Integrating Gemini's advanced reasoning with Veo's video generation capability, it enables native "any-to-any" generation, producing high-quality 24 FPS videos with synchronized audio directly from text prompts. It supports 16:9 and 9:16 aspect ratios, 3–10s video duration, and up to 1080P/4K supersampled output. Featuring multi-turn conversational video editing with context retention and built-in SynthID digital watermarking, it is ideal for short-form video creation, commercial ads, film VFX, and multimodal Agent workflows.
Gemini Omni Flash Image-to-Video
Gemini Omni Flash Image-to-Video is a next-generation multimodal video generation model developed by Google DeepMind. Built on a native Omni architecture, it accurately parses text prompts and input image semantics to produce cinematic 24 FPS dynamic videos. Supporting 16:9 and 9:16 aspect ratios, it generates 3–10 second fluid clips per run (featuring native support for up to 4K super-sampled upscaling, with 720P currently available on select platform endpoints). With exceptional subject consistency, physical simulation, and camera control, it excels in short-form drama, commercial advertising, film VFX, and social media animation.
Gemini 3.6 Flash
Gemini 3.6 Flash is Google's next-generation lightweight workhorse model released in July 2026. Built for agentic workflows, complex coding, and multimodal tasks, it supports a 1M token input context window and a 64K token output limit. Compared to 3.5 Flash, it reduces output token consumption by ~17% with streamlined reasoning steps and tool calls, significantly lowering overall cost and latency for agent execution. Natively handling text, image, video, audio, and PDF inputs, it excels at computer use and multi-tool orchestration.
Read Me
Gemini 3.8 Flash
Gemini 3.8 Flash is a multimodal large language model released by Google DeepMind in September 2026, building on Gemini 3.7 Flash with performance advancements across software engineering and agentic knowledge workflows. The model accepts text, image, audio, video, and PDF input with text output, a 1 million token context window, 64K max output, and a knowledge cutoff of March 2026. It supports three thinking levels (low/medium/high, default medium) and improves across Terminal-Bench 2.1 (90.8%), SWE-Bench Pro (61.6%), and CharXiv (86.2%) compared to 3.7 Flash.
The iCreat platform offers this model via the endpoint https://api.icreat.ai/llm/openai/v1/chat/completions (OpenAI Chat Completions compatible), at $1.5 per million input tokens, $7.5 per million output tokens, and $0.15 per million cached read tokens.
Model Positioning
Gemini 3.8 Flash is positioned as Google DeepMind's efficient, lightweight multimodal workhorse model, designed for high-concurrency, low-latency agentic workflows, code building, and multimodal understanding.
Building on Gemini 3.7 Flash, the model delivers significant performance improvements in software engineering and agentic knowledge workflows while maintaining extreme inference speed and high cost efficiency. Customizable thinking levels (low/medium/high) let developers balance quality, cost, and latency.
On the iCreat platform, the model is served through an OpenAI Chat Completions compatible endpoint, and existing OpenAI SDKs work directly.
Core Capabilities
Omni-Modal Input
Supports text, image, audio, video, and PDF input with text output (64K max tokens), covering document understanding, chart analysis, video comprehension, and audio processing multimodal scenarios.
1 Million Token Ultra-Long Context
Supports 1,048,576 token context window, enabling processing of large codebases, complete document sets, and long-running agent histories in a single call.
Customizable Thinking Levels
Supports three thinking levels (low/medium/high, default medium), allowing developers to flexibly balance reasoning depth against cost and latency. Lower levels reduce cost and latency; higher levels improve reasoning depth.
Coding and Agent Performance
Terminal-Bench 2.1 at 90.8% (up from 3.7 Flash's 81.6%), SWE-Bench Pro at 61.6%, SWE-Atlas at 51.9% — leading in coding and agentic terminal tasks.
Streaming Output and Function Calling
Supports streaming output (SSE), function calling, structured output, code execution, prompt caching, and batch processing across the Chat Completions endpoint.
Pricing
| Token Type | Unit Price | Unit |
|---|---|---|
| Input | $1.5 | per million tokens |
| Output | $7.5 | per million tokens |
| Cached Read | $0.15 | per million tokens |
| Cache Write | $0.08 | per million tokens |
Note: The above are iCreat platform prices. Google's official pricing is $1.50 input, $7.50 output per million tokens (promotional rate $0.75/$3.75 through 2026-12-31), cache read $0.15. The iCreat platform price shall prevail.
Application Scenarios
- Agentic workflows and multi-step tool calling
- Software engineering and coding tasks (multi-file development, debugging, refactoring)
- Long-document analysis and large-scale codebase understanding
- Multimodal document processing (charts, PDFs, video, audio)
- Enterprise API integration and real-time interaction scenarios
Model Comparison
Comparison Table 1: Gemini 3.8 Flash vs 3.7 Flash
| Feature | Gemini 3.8 Flash | Gemini 3.7 Flash |
|---|---|---|
| Positioning | 3.8 iterative upgrade | 3.7 base version |
| Release Date | 2026-09 | 2026-08-13 |
| Context Window | 1M tokens | 1M tokens |
| Max Output | 64K tokens | 64K tokens |
| Knowledge Cutoff | 2026-03 | 2026-03 |
| Terminal-Bench 2.1 | 90.8% | 81.6% |
| SWE-Bench Pro | 61.6% | 60.4% |
| SWE-Atlas | 51.9% | 48.0% |
| CharXiv | 86.2% | 84.5% |
| GDP.pdf | 35.0% | 34.0% |
| Input Price | $1.5/M tokens | $0.75/M tokens |
| Output Price | $7.5/M tokens | $3.75/M tokens |
Note: 3.8 Flash improves across coding, agent, and multimodal benchmarks over 3.7 Flash, priced at 2x.
Comparison Table 2: Same-Tier Multimodal Models
| Feature | Gemini 3.8 Flash | Claude Sonnet 5 | GPT-5.6 Terra |
|---|---|---|---|
| Positioning | Multimodal workhorse | Enterprise coding model | General flagship |
| Input Price | $1.5/M tokens | $2.00/M tokens | $2.00/M tokens |
| Output Price | $7.5/M tokens | $10.00/M tokens | $12.00/M tokens |
| Context Window | 1M tokens | 1M tokens | 1M tokens |
| Input Modalities | Text+Image+Audio+Video+PDF | Text+Image | Text+Image |
| Terminal-Bench 2.1 | 90.8% | Per official | Per official |
| Max Output | 64K tokens | 128K tokens | 128K tokens |
Note: Gemini 3.8 Flash is the lowest-priced of the three with the broadest input modality support (only one supporting audio+video+PDF) and top coding score; max output is smaller than Claude/GPT (64K vs 128K).
Why Choose Gemini 3.8 Flash?
- Omni-modal input (text+image+audio+video+PDF), covering the broadest multimodal scenarios
- 1 million token ultra-long context for large codebases and long documents in a single call
- Terminal-Bench 2.1 at 90.8%, leading coding and agent performance
- Customizable thinking levels (low/medium/high) for flexible quality-cost balance
- Lowest price among comparable models ($1.5/$7.5), far below Sonnet 5 and GPT-5.6 Terra
- OpenAI Chat Completions compatible endpoint — existing SDKs work directly
Specifications
| Field | Value |
|---|---|
| Model Name | Gemini 3.8 Flash |
| Developer | Google DeepMind |
| Model ID | gemini-3.8-flash |
| Endpoint | https://api.icreat.ai/llm/openai/v1/chat/completions |
| SDK base_url | https://api.icreat.ai/llm/openai/v1 |
| Release Date | 2026-09 |
| Model Type | Multimodal LLM |
| Authentication | Authorization: Bearer |
| Context Window | 1,048,576 tokens (1M) |
| Max Output | 65,536 tokens (64K) |
| Knowledge Cutoff | 2026-03 |
| Input Modalities | Text, Image, Audio, Video, PDF |
| Output Modalities | Text |
| Thinking Level | low / medium (default) / high |
| Streaming | Supported |
| Function Calling | Supported |
| Structured Output | Supported |
| Code Execution | Supported |
| Prompt Caching | Supported |
| Batch Processing | Supported |
| Billing Unit | Per million tokens |
Architecture
The iCreat platform's Gemini 3.8 Flash is served through an OpenAI Chat Completions compatible endpoint, allowing users to call it with any OpenAI API-compatible SDK. Requests must include the API Key in the Authorization header. The request body contains the model field (gemini-3.8-flash), a messages array (supporting multimodal content: text, image_url, video_url, input_audio), and a modalities field (specifying output modality). The model is based on Gemini 3.7 Flash with iterative upgrades, supporting customizable thinking levels (low/medium/high) and delivering significant performance improvements in software engineering and agentic knowledge workflows. The model supports streaming (stream: true), function calling, structured output, code execution, and prompt caching.
Notes
- The model supports omni-modal input (text+image+audio+video+PDF) but output is text only
- Max output is 64K tokens, smaller than Claude Opus 5 (128K) and Fable 5.1 (128K)
- Thinking level defaults to medium; adjust to low or high based on task complexity
- Cache read costs only $0.15/M tokens — one-tenth of the standard input rate, significantly reducing costs for repeated queries
- API Key should be passed via environment variables; avoid hardcoding it in source code
- Please safeguard your API Key and avoid hardcoding it in client-side code or public repositories
Frequently Asked Questions
What is the difference between Gemini 3.8 Flash and 3.7 Flash?
3.8 Flash builds on 3.7 Flash with iterative upgrades, improving across coding (Terminal-Bench 2.1 90.8% vs 81.6%), software engineering (SWE-Bench Pro 61.6% vs 60.4%), and multimodal (CharXiv 86.2% vs 84.5%), priced at 2x.
What input modalities are supported?
Supports text, image, audio, video, and PDF — five input modalities. Output is text only (64K max tokens). In the content array, specify input type via the type field: text, image_url, video_url, input_audio.
How do I control the thinking level?
Three levels: low, medium (default), and high. Lower levels reduce cost and latency; higher levels improve reasoning depth. Use medium for everyday tasks, high for complex reasoning, and low for rapid response.
How large is the context window?
Supports 1 million tokens (1,048,576 tokens) context window with 64K (65,536 tokens) max output. Can process large codebases, complete document sets, and long-running agent histories in a single call.
Is streaming output supported?
Yes. Pass "stream": true in the request body to enable streaming. The model will push incremental content via Server-Sent Events (SSE), suitable for real-time display.
Is function calling supported?
Yes. The model is compatible with the OpenAI Function Calling protocol, and also supports structured output (JSON Schema), code execution, prompt caching, and batch processing. Native web search is not supported (use external search tools via function calling).