What is Gemini 3.7 Flash?
Gemini 3.7 Flash is Google's newest Gemini model, released August 13, 2026, three weeks after Gemini 3.6 Flash. It is the default fast model of the lineup: multimodal input across text, image, audio, video, and PDF, a 1M-token context window, 64k output tokens, and output speed around 340 tokens per second. It runs on Unifically as google/gemini-3.7-flash. Thinking is tunable in three levels (low, medium, high), and the knowledge cutoff reaches March 2026 for code, SDKs, and technical domains.
What's new in Gemini 3.7 Flash
More intelligence at Flash speed
Artificial Analysis scores it 56 on the Intelligence Index at high thinking, one point behind GPT 5.6 Terra Max, while outputting around 340 tokens per second at every thinking level. Even the medium setting scores 53, one point above Gemini 3.6 Flash at high.
The best long-context recall in the Flash line
97.0% on GDM-MRCR v2 at 128k tokens, ahead of GPT 5.6 Terra at 93.5% and Claude Sonnet 5 at 81.5%, and 62.5% on the 1M pointwise variant. Long documents stay retrievable deep into the window.
Coding that edges out bigger names
On FrontierCode 1.1 Main it scores 43.6%, just ahead of Claude Sonnet 5 at 42.7% and GPT 5.6 Terra at 41.3%, and far above Gemini 3.6 Flash at 34.4%. On the WebDev Arena board its preliminary 1588 Elo is 50 points up on its predecessor.
Full multimodal input
Text, image, audio, video, and PDF go in; text comes out. One model covers transcription-adjacent work, chart reading, video Q&A, and document extraction.
Best for
High-volume agent steps
Around 340 tok/s output keeps multi-step chains moving without queue buildup.
Web dev and UI generation
Preliminary 1588 Elo on WebDev Arena, 50 points above Gemini 3.6 Flash.
Long document work
97% recall at 128k tokens; contracts, codebases, and research packs stay addressable.
Video and audio understanding
Native video, audio, and PDF input without a separate transcription step.
Cost-tiered reasoning
Drop to low thinking for cheap fast calls, raise to high for hard problems.
Document extraction pipelines
Strong PDF handling and structured output for claims, tables, and fields.
Use cases
Use Gemini 3.7 Flash for agent fleets where each step needs decent reasoning but latency budgets are tight: support triage, research summarization, and workflow automation. It fits web-dev copilots that scaffold and revise UIs across many quick turns. Feed it hour-long meeting recordings or lecture video and get structured notes back in one call. And with 1M tokens of context plus 97% retrieval at 128k, it handles whole-repo questions and long contract review without aggressive chunking.
Limitations
Output is text only, and thinking tokens are billed as output tokens even though only summaries are returned.
Most headline agentic numbers (FrontierCode, DeepSWE, GDM-MRCR) come from Google's own evaluation run; the independent picture is thinner, and GPT 5.6 Terra stays ahead on DeepSWE v1.1 (69.6% against 65.3%) and Terminal-Bench 2.1 (87.4% against 85.8%) in the same table.
Its Arena text score (1490 Elo, rank #9) is marked preliminary, and the minimal thinking level from earlier Flash models is gone, so the cheapest classification-style calls now carry some thinking cost.
Gemini 3.7 Flash vs Gemini 3.6 Flash
The upgrade is real across the board: FrontierCode 1.1 rises from 34.4% to 43.6%, DeepSWE v1.1 from 48.6% to 65.3%, 128k MRCR from 91.8% to 97.0%, and WebDev Arena from 1538 to 1588 Elo. Speed goes up too, roughly 340 tokens per second against 234 for the older model. Unless a pipeline depends on regression-tested Gemini 3.6 Flash behavior, new work belongs on 3.7 Flash.
When to use Gemini 3.7 Flash
Use Gemini 3.7 Flash as the default fast model for agents, web dev, and multimodal document work when you want near-frontier answers at Flash latency. Step up to a frontier model when the task is deep single-shot reasoning; step down to a smaller model only for pure classification and routing.
API examples
Call Gemini 3.7 Flash from any language by POSTing to /v1/chat/completions, the OpenAI-compatible endpoint shared by every language model on the platform. Full parameter docs live at docs.unifically.com/models/llm/google/gemini-3.7-flash.
curl -X POST https://api.unifically.com/v1/chat/completions \
-H "Content-Type: application/json" \
-H "Authorization: Bearer YOUR_API_KEY" \
-d '{
"model": "google/gemini-3.7-flash",
"messages": [
{ "role": "user", "content": "Summarize the attached contract and list every obligation with a deadline." }
]
}'
The response comes back synchronously with the completion. Set "stream": true to receive tokens as they generate.
FAQs
People also ask
google/gemini-3.7-flash, called through the OpenAI-compatible POST /v1/chat/completions endpoint with your Unifically API key.
Text, image, audio, video, and PDF input, with text output up to 64k tokens per response. The context window is 1M tokens.
Yes, across Google's evaluation suite. FrontierCode 1.1 rises from 34.4% to 43.6%, DeepSWE v1.1 from 48.6% to 65.3%, GDM-MRCR v2 at 128k from 91.8% to 97.0%, and its Artificial Analysis Intelligence Index is 56 against 52.
Artificial Analysis measures around 340 output tokens per second at every thinking level, which is faster than most frontier models. Time to first token stays under a second at low thinking.
Three, low, medium, and high, with medium as the default. Higher levels reason longer before answering; thinking tokens are billed as output tokens.
No. As of August 2026 the Pro line is still at Gemini 3.1, and Google has said the next Pro release is delayed. Gemini 3.7 Flash is the newest Gemini model overall.
97.0% on GDM-MRCR v2 at 128k tokens and 62.5% on the 1M pointwise variant, the best long-context retrieval scores in the Flash line.
