What it is
Gemini 3.8 Flash is Google's main mid-priced model, released on 2 September 2026. Google calls it its most intelligent Flash model, built for long-running software engineering, agents and business workflows. It reads text, images, video, audio and PDFs, writes text, and has a 1,048,576-token input limit with 65,536 output tokens. Since 8 October, API calls to gemini-3.7-flash are routed to it automatically.
API pricing is an introductory $0.75 per million input tokens and $3.75 per million output tokens until 31 December 2026, then $1.50/$7.50. Batch and Flex halve those rates, and it is on the Gemini API free tier with rate limits.
It is in the Gemini API, Google AI Studio, Android Studio, Antigravity, Gemini Enterprise, Google Search's AI Mode, Google Sheets, and the Gemini app for AI Pro and Ultra subscribers.
Our review
Gemini 3.8 Flash looked like the value pick at launch and looks much less so on today's independent numbers. Artificial Analysis now scores it 41 on its Intelligence Index at high reasoning, level with Claude Sonnet 5.5 at medium effort and below Claude Haiku 5.5 at max (43). Its cost per index task is $1.24, against $0.21 for Haiku 5.5 and $0.32 for GPT-6.1 Sol at high effort, which scores 50.
Verbosity explains the gap. Artificial Analysis counts 170M output tokens to run its index, nearly double the median of 87M. Google warned of this at launch, saying the model "works harder" and may use more tokens at higher effort, and The Register reported cost per task already 40% above 3.7 Flash at similar per-token prices. Low list prices help little when the model writes twice as much.
Input breadth is where it still earns its place. It takes audio, video and PDFs directly alongside text and images, with Search and Maps grounding, code execution and a free API tier; the Claude models read only text and images. For a team that needs to transcribe a call, read a scanned form and answer over both, one request replaces a pipeline.
Two operational details matter. Artificial Analysis measures 24 seconds to the first token on Google's API, which rules it out for chat where a user is waiting. And on 1 January 2027 the price doubles to $1.50/$7.50, three-quarters of Sonnet 5.5's list price for a model that scores like Sonnet at medium effort.
Where it earns its place
An insurance claims team processing 20,000 claim files a month
Read photos, PDFs and recorded calls and draft a summary
Send each claim's files in one request within the 1M-token input, with no separate transcription or OCR step. The Batch API halves the cost to $0.375/$1.875 until January; budget on the doubled rate after that.
A product team prototyping a multimodal feature
Test video and audio understanding before committing budget
The free tier covers standard input, output and caching at no cost, but Google may have humans review free-tier prompts and use them to improve its products, so stick to synthetic or public data until billing is on.
A platform engineer choosing a model for a 1M-replies-a-month support bot
Find the cheapest model that answers well
At 2,000 input and 300 output tokens a reply, list price is about $2,625 a month now and $5,250 from January, before thinking tokens, against about $350 for Claude Haiku 5.5. The 24-second first-token latency is a further reason to look elsewhere for live chat.
Before you commit
- Setup
- An AI Studio key and model ID gemini-3.8-flash; thinking level can be low, medium or high, and minimal returns an error.
- Integrations
- Gemini API and SDKs; Batch, Flex and Priority inference; function calling, structured outputs, code execution, URL context and file search; computer use in preview. No Live API.
- Data and privacy
- Paid-tier prompts are not used to improve Google's products and are kept for a limited, unspecified period for abuse checks; free-tier prompts may be read by human reviewers and used for training. Grounded requests are stored for 30 days.
- Cost at scale
- $0.75/$3.75 per million tokens until 31 December, then $1.50/$7.50; cache reads $0.075 rising to $0.15, plus $0.50 per million tokens per hour of cache storage. Thinking is billed as output, which matters for a model this verbose.
- Limits
- 1,048,576 input and 65,536 output tokens. Per-model rate limits are shown only in AI Studio; Tier 1 accounts have a $250 billing cap, a $10 spend limit per rolling 10 minutes and 3M queued batch tokens.
What's good
- Reads text, images, audio, video and PDFs in one call
- 1,048,576-token input with Search and Maps grounding
- Free API tier for prototyping
- 141.5 output tokens a second once it starts
What's not
- $1.24 per Artificial Analysis index task, against $0.21 for Haiku 5.5
- Very verbose: 170M output tokens to run the index
- 24-second time to first token on Google's API
- Price doubles on 1 January 2027
- Free-tier prompts may be used by Google
Best for
Teams processing audio, video and long PDFs inside Google's ecosystem who can tolerate slow first responses.
Alternatives
Gemini 3.8 Flash: common questions
- Is Gemini 3.8 Flash worth it?
- Broad inputs and a free tier keep Gemini 3.8 Flash useful, but it now spends more per task than rivals that score higher, and its price doubles in January. We rate it 4.1 out of 5.
- How much does Gemini 3.8 Flash cost?
- API: $0.75 input / $3.75 output per 1M tokens until 31 Dec 2026, then $1.50 / $7.50; Batch and Flex half price; free tier with limits.
- What are the downsides of Gemini 3.8 Flash?
- $1.24 per Artificial Analysis index task, against $0.21 for Haiku 5.5; very verbose: 170M output tokens to run the index; 24-second time to first token on Google's API; price doubles on 1 January 2027; free-tier prompts may be used by Google.
- Who is Gemini 3.8 Flash best for?
- Teams processing audio, video and long PDFs inside Google's ecosystem who can tolerate slow first responses.
- How much does Gemini 3.8 Flash cost for a team or in production?
- $0.75/$3.75 per million tokens until 31 December, then $1.50/$7.50; cache reads $0.075 rising to $0.15, plus $0.50 per million tokens per hour of cache storage. Thinking is billed as output, which matters for a model this verbose.
- Is Gemini 3.8 Flash safe for company data?
- Paid-tier prompts are not used to improve Google's products and are kept for a limited, unspecified period for abuse checks; free-tier prompts may be read by human reviewers and used for training. Grounded requests are stored for 30 days.
- What are the best Gemini 3.8 Flash alternatives?
- The closest alternatives we have reviewed are Claude Sonnet 5.5, GPT-6.1 Sol, Gemini 3.5 Flash-Lite and Gemini 4 Argon.
Answers checked Oct 22, 2026.
Make Gemini 3.8 Flash? Show this rating on your site
