- Gemini 3.7 Flash landed on August 13, 2026, only three weeks after 3.6 Flash.
- Introductory pricing is $0.75 in / $3.75 out per 1M tokens: half of 3.6 Flash’s standard rate. The promo ends December 31, 2026.
- Coding jumped hard: FrontierCode 1.1 went from 34.4% to 43.6%, DeepSWE v1.1 from 48.6% to 65.3%.
- Context window is 1,048,576 tokens with a 65,536 token output cap. It takes text, images, video, audio and PDF.
- It is not the smartest model on the board: 56 on the Artificial Analysis index, one point behind GPT-5.6 Terra and Muse Spark 1.2, at roughly a third of their price.
Google’s Flash line has never been about winning the intelligence crown. It is the model you run at volume: cheap enough to call thousands of times a day, fast enough to sit inside a product. Gemini 3.7 Flash stretches that definition.
Shipped three weeks after Gemini 3.6 Flash, it posts a clear jump in coding and agent work while cutting the price in half. Google’s own framing calls it “our most intelligent workhorse model yet for coding and agents.”
The company credits the fast turnaround to developer feedback plus algorithmic work it expects to carry into future models. The announcement came from Tulsee Doshi, Senior Director of Product Management, on behalf of the Gemini team.
Let’s look at the numbers. 👇🏻
What Is Gemini 3.7 Flash?
Gemini 3.7 Flash is the newest member of Google’s Flash family. The split is simple: Pro models handle the hardest single problems, Flash models handle the volume. If you are firing thousands of requests a day or running an agent for hours, this is the model that decides your bill.
The spec sheet:
| Property | Value |
|---|---|
| Model id | gemini-3.7-flash |
| Context window | 1,048,576 tokens |
| Output limit | 65,536 tokens |
| Input | Text, image, video, audio, PDF |
| Output | Text only |
| Thinking levels | low, medium, high |
| Open weights | None |
One detail that will bite people porting code over: the minimal thinking level is not supported here and returns an error. If your 3.6 Flash config sets it, change it before you ship.
The model supports caching, code execution, file search, function calling, grounding with Google Maps and Search, structured outputs and URL context. Computer use is in preview. Image generation, audio generation and the Live API are not available.
How much is 1M tokens?
Roughly 750,000 words: an entire mid-sized codebase, or a few hundred pages of technical documentation. You can measure your own text in seconds with our token counter.
What Changed in This Release
Google highlights three areas.
Software engineering. Debugging and issue resolution improved noticeably over 3.6 Flash, and first-pass code accuracy went up. FrontierCode 1.1 Main, which measures production-ready code quality, moved from 34.4% to 43.6%. DeepSWE v1.1, a long-horizon software engineering eval, went from 48.6% to 65.3%.
Web development. The model produces more functional layouts and more feature-complete apps in fewer prompts. For UI generation it holds high design adherence against a reference input, whether that is a screenshot, an image or a full design system. On Arena.ai’s Code Arena it moved from 1538 to 1588 Elo.
Knowledge work. In dense fields like finance, law and biosciences, reasoning and accuracy improved. On GDP.pdf, a complex-document comprehension eval, it went from 22.0% to 34.0%. On AutomationBench, which measures real-world business workflow automation, from 17.0% to 30.4%.
On developer experience, the behavioural change Google leans on is this: the model adapts better to roadblocks, asks for clarification when intent is ambiguous, and follows instructions more faithfully. It puts more effort into multi-step planning and tool calls. In practice that means less manual oversight and fewer retries.
Benchmark Results 📊
Here is Google’s published comparison, against Claude Sonnet 5, GPT-5.6 Terra and Muse Spark 1.2.
| Benchmark | 3.7 Flash | 3.6 Flash | Sonnet 5 | GPT-5.6 Terra | Muse Spark 1.2 |
|---|---|---|---|---|---|
| Input price ($/1M) | 0.75* | 0.75* | 2.00 | 2.00 | 1.25 |
| Output price ($/1M) | 3.75* | 3.75* | 10.00 | 12.00 | 4.25 |
| AA Intelligence Index | 56 | 52 | 55 | 57 | 57 |
| FrontierCode 1.1 Main | 43.6% | 34.4% | 42.7% | 41.3% | - |
| DeepSWE v1.1 | 65.3% | 48.6% | 53.8% | 69.6% | 54.9% |
| Code Arena (Elo) | 1588 | 1538 | 1541 | 1523 | 1535 |
| Terminal-bench 2.1 | 85.8% | 78.0% | 80.4% | 87.4% | 82.9% |
| Terminal-bench 3.0 | 14.9% | 5.4% | 14.6% | 20.8% | - |
| AutomationBench | 30.4% | 17.0% | 10.7% | 23.6% | - |
| GDPVal-AA v2 (Elo) | 1525 | 1422 | 1598 | 1578 | 1628 |
| Harvey LAB-AA | 90.7% | 85.1% | 90.1% | 85.2% | - |
| GDP.pdf | 34.0% | 22.0% | 28.0% | 24.7% | 16.0% |
| LVBench | 85.4% | 84.2% | 68.5% | 78.9% | - |
| GDM-MRCR v2 (128k) | 97.0% | 91.8% | 81.5% | 93.5% | - |
| OSWorld-2.0 | 47.9% | 33.8% | - | 50.2% | - |
| HLE-Verified | 53.6% | 51.2% | 31.0% | 51.1% | - |
| LABBench2 | 82.1% | 76.1% | 80.1% | 81.2% | - |
The summary: 3.7 Flash beats 3.6 Flash on every row, and against the competition the wins and losses split cleanly.
It leads on production code quality (FrontierCode), web development (Code Arena), enterprise workflow automation (AutomationBench), document comprehension (GDP.pdf), long video understanding (LVBench) and long-context retrieval (GDM-MRCR). That last one, at 97.0%, is the number that makes the 1M token window worth quoting.
It trails on pure agentic and long-horizon engineering: GPT-5.6 Terra is four points ahead on DeepSWE, six on Terminal-bench 3.0 and two and a half on OSWorld-2.0. On the knowledge-work eval GDPVal-AA v2 it sits behind all three rivals.
On Artificial Analysis’ intelligence index the model scores 56: four points above 3.6 Flash (52), one above Claude Sonnet 5 (55), one below GPT-5.6 Terra and Muse Spark 1.2 (57).
Pricing and the December 31 Catch 💸
Per 1M tokens:
| Period | Input | Output |
|---|---|---|
| Introductory (through Dec 31, 2026) | $0.75 | $3.75 |
| From January 1, 2027 | $1.50 | $7.50 |
The price you see today is not the price you keep. The rate doubles at the start of the year, which puts the model exactly where 3.6 Flash’s standard pricing sat. That is also where the “half the cost” claim comes from: the comparison is against 3.6 Flash’s original price, not its promo.
If you are budgeting, model both scenarios. Your cost doubles overnight in January.
Want your own numbers? Drop your input and output token counts into our LLM cost calculator and compare Gemini against GPT and Claude side by side.
From a Static PDF to an Interactive Report
Document handling is where the gain is most visible. The GDP.pdf jump from 22.0% to 34.0% shows up in exactly the material you would expect to be painful: annual reports stuffed with tables, footnotes and charts.
Legal work shows the same pattern: on Harvey LAB-AA, which covers complex legal workflows, 90.7% is the highest score in the table. On LABBench2, covering real-world biology research tasks, 82.1% again puts it ahead of its rivals.
Gemini Spark Now Runs on 3.7 Flash
Gemini Spark, available to Google AI Pro and Ultra subscribers in more than 160 countries, switches to 3.7 Flash today. Spark launched at I/O as a personal agent that runs 24/7 and takes action under your direction: consolidating files, drafting emails, updating status documents.
With this update, tool use across Google Workspace apps improves, and accuracy and output quality go up on complex, multi-skill workflows.
Some regions are out of scope: the European Economic Area, Nigeria, Switzerland and the United Kingdom are excluded. Google did not publish the full country list, so check your own account.
Safety
Google says it extended its Frontier Safety coverage for this release. 3.7 Flash ships with updated safeguards against misuse in Chemical, Biological, Radiological and Nuclear (CBRN) and cyber offense domains, while aiming not to block legitimate use, such as defensive security work like vulnerability patching.
Where You Can Use It
As of August 13, 2026:
- Developers: agent-first workflows in Google Antigravity, or the Gemini API via Google AI Studio and Android Studio.
- Enterprises: Gemini Enterprise Agent Platform and the Gemini Enterprise app.
- Individuals: the Spark agent inside the Gemini app, for AI Pro and Ultra subscribers.
There are no open weights, so you cannot download it and run it on your own hardware.
For Developers: Five Lines to Start
from google import genai
from google.genai import types
client = genai.Client(api_key="YOUR_KEY")
response = client.models.generate_content(
model="gemini-3.7-flash",
contents="Explain Gemini 3.7 Flash in one sentence.",
config=types.GenerateContentConfig(
thinking_config=types.ThinkingConfig(thinking_level="low"), # low | medium | high
),
)
print(response.text)
Three things worth knowing:
- No
minimal. Code carried over from 3.6 Flash that sets it will error. Uselowinstead. - Use caching. If you resend the same system prompt or the same file repeatedly, caching cuts the bill substantially.
- Thinking level is billed. Tokens the model produces while thinking count as output, and output costs five times input. For simple classification,
lowis enough.
Which Model for Which Job?
The benchmark split turns into a fairly clean recommendation:
- Long documents, PDFs, reports, archives: 3.7 Flash, clearly. Its GDP.pdf and GDM-MRCR scores beat every rival in the table, and it actually uses that 1M token window rather than degrading inside it.
- Web UI and fast prototyping: 3.7 Flash again. It holds the highest Code Arena Elo on the board, which is where the single-shot interface claim comes from.
- Autonomous agents running for hours, terminal work: GPT-5.6 Terra still leads. The DeepSWE and Terminal-bench 3.0 gaps are not noise, so if budget allows, run the critical agent there.
- High-volume production workloads: this is Flash’s home turf. On a classification, summarisation or support pipeline chewing through millions of tokens a day, a third of the price matters far more than one point of index score.
A mixed setup is the obvious answer: a router that sends cheap work to Flash and hard agent steps to a pricier model is the most common cost optimisation running in production today.
What Comes Next
The cadence is the notable part: 3.6 Flash on July 21, 3.7 Flash on August 13. Google explicitly describes the work behind this release as algorithmic innovations it looks forward to bringing to future models, which reads like an interim step rather than a destination.
A Pro-tier refresh, or a response from competitors, is unlikely to be far off. Flash holding this price-per-point band puts the pressure squarely on the cheap end of the market.
Frequently Asked Questions
Q: When was Gemini 3.7 Flash released? A: Google announced it on August 13, 2026 and made it available the same day through the Gemini API, Google AI Studio, Android Studio, Antigravity and Gemini Enterprise.
Q: How much does Gemini 3.7 Flash cost? A: $0.75 per 1M input tokens and $3.75 per 1M output tokens as an introductory price. That ends December 31, 2026; from January 1, 2027 it becomes $1.50 and $7.50.
Q: Is Gemini 3.7 Flash free? A: API usage is paid. Google AI Studio offers limited free trial usage. The Spark agent in the Gemini app requires a Google AI Pro or Ultra subscription.
Q: Is Gemini 3.7 Flash better than GPT-5.6 or Claude? A: Mixed. It scores 56 on the Artificial Analysis index, above Claude Sonnet 5 (55) and below GPT-5.6 Terra (57). It leads on FrontierCode, Code Arena, GDP.pdf and AutomationBench, and trails on DeepSWE, Terminal-bench 3.0 and GDPVal. On price per point it is clearly ahead.
Q: What is the difference between Gemini 3.7 Flash and 3.6 Flash? A: 3.7 Flash wins every benchmark. The biggest gaps are DeepSWE (48.6% → 65.3%), FrontierCode (34.4% → 43.6%), GDP.pdf (22.0% → 34.0%) and AutomationBench (17.0% → 30.4%). The minimal thinking level was also dropped.
Q: What is the Gemini 3.7 Flash context window? A: 1,048,576 input tokens with a 65,536 token output cap. It scored 97.0% on the GDM-MRCR v2 long-context eval at a 128k average.
Take care… 🙂