Gemini 3.7 Flash Is Live: Big Coding and Agent Gains

Key Takeaways

The new Flash workhorse: Google released gemini-3.7-flash on August 13, 2026, positioning it as its “most intelligent workhorse model” for coding and agents. It is now live on APIYI
Unusually large coding gains: DeepSWE v1.1 goes from 48.6% → 65.3%, FrontierCode 1.1 from 34.4% → 43.6%, Terminal-bench 2.1 from 78.0% → 85.8%
Business automation nearly doubles: AutomationBench moves from 17.0% → 30.4%, ahead of Claude Sonnet 5 (10.7%) and GPT-5.6 Terra (23.6%) in Google’s own comparison
Same price as Google’s official rate: $0.75 in / $3.75 out per 1M tokens — this is Google’s limited-time promotional price, valid through December 31, 2026; standard pricing of $1.50 / $7.50 resumes January 1, 2027
Specs unchanged: 1M context / 64K output, multimodal input (text, image, audio, video), tunable thinking levels (low / medium / high)

Background

Flash is the mid-tier workhorse of Google’s Gemini family — Pro sits above it, Flash-Lite below for high-frequency lightweight work, and Flash covers the vast majority of production workloads where capability has to be good enough and the price has to survive real volume. This tier has shipped several versions in the past six months (3.5 → 3.6 → 3.7), making it the fastest-moving line in Gemini and our standing default recommendation. 3.7 Flash arrives roughly three weeks after 3.6 Flash. Google put the weight squarely on coding and agents: not just writing code, but producing deployable, production-ready code on the first try, and actually finishing long-horizon, multi-step automation tasks. All figures here come from the official Google DeepMind model card and the Gemini API documentation, collected August 14, 2026.

Deep Dive

Official benchmarks (3.7 Flash vs 3.6 Flash)

Benchmark	3.7 Flash	3.6 Flash	What it measures
DeepSWE v1.1	65.3%	48.6%	Long-horizon software engineering
FrontierCode 1.1	43.6%	34.4%	Production code quality
Terminal-bench 2.1	85.8%	78.0%	Terminal-environment tasks
Terminal-bench 3.0	14.9%	5.4%	Harder terminal tasks
Code Arena Web (Elo)	1588	1538	Web development
AutomationBench	30.4%	17.0%	Real business workflow automation
OSWorld-2.0	47.9%	33.8%	GUI operation
GDM-MRCR v2	97.0%	91.8%	Long-context retrieval (128k)
GDP.pdf	34.0%	22.0%	PDF document comprehension
HLE-Verified	53.6%	51.2%	Hard knowledge reasoning
CharXiv Reasoning (no tools)	84.5%	85.2%	Chart reasoning — slight regression

The only regression is CharXiv chart reasoning (84.5% vs 85.2%; 88.7% vs 89.4% with tools), a gap within noise. If chart-heavy analysis is your primary workload, run your own A/B before switching.

Technical specs

Item	Value
Model ID	`gemini-3.7-flash`
Context window	1,000,000 tokens
Max output	64,000 tokens
Input modalities	Text / image / audio / video
Output modality	Text
Knowledge cutoff	March 2026 (January 2025 for some domains)
Thinking levels	low / medium (default) / high
Release date	August 13, 2026

The thinking levels are straightforward: low cuts latency for time-sensitive paths, medium is the default and the quality/cost balance point for everyday work, and high is reserved for complex reasoning and hard coding tasks.

Google states that 3.7 Flash carries the same built-in tool suite as 3.6 Flash (code execution, Google Search, Maps, etc.), with structured outputs and multimodal support unchanged. An item-by-item APIYI test report will follow; if you depend heavily on one specific tool, validate on low traffic first.

In Practice

Recommended use cases

Coding and code review: the biggest gain in this release, especially where deployable code on the first pass matters
Agents / workflow automation: AutomationBench nearly doubled — multi-step completion rate is the headline of this version
Long documents and long context: GDM-MRCR v2 at 97.0% means near-lossless retrieval across the 1M window, with a clear jump in PDF comprehension
Everyday primary chat and batch jobs: the mid-tier position, where capability meets unit price

Code examples

from openai import OpenAI

client = OpenAI(api_key="YOUR_API_KEY", base_url="https://api.apiyi.com/v1")

response = client.chat.completions.create(
    model="gemini-3.7-flash",
    messages=[
        {"role": "user", "content": "Write a Python HTTP client with retries and exponential backoff"}
    ]
)
print(response.choices[0].message.content)

curl -X POST "https://api.apiyi.com/v1beta/models/gemini-3.7-flash:generateContent" \
  -H "x-goog-api-key: YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "contents": [{"parts": [{"text": "Write a Python HTTP client with retries and exponential backoff"}]}]
  }'

The native endpoint takes your APIYI token directly (x-goog-api-key: sk-...) — no Google API Key needed.

Best practices

Migrating from 3.6 Flash: change the model name, nothing else — the request shape is identical, and during the limited-time promotion the unit price is half
Cache repeated prefixes: system prompts and long fixed documents read back at $0.075 / 1M tokens, one tenth of the input price
Match thinking level to task: low for batch classification and extraction, high for complex refactors and multi-step agents — don’t set one level globally

Pricing & Availability

Billing item	Price (per 1M tokens)
Input (prompt)	$0.7500
Output (completion, incl. thinking)	$3.7500
Cache read	$0.0750
Cache write (5m)	$0.7500

Identical to Google’s official pricing. Worth spelling out: $0.75 / $3.75 is Google’s own limited-time promotional rate, officially valid through December 31, 2026, reverting to $1.50 / $7.50 on January 1, 2027 (the same price as 3.6 Flash). In other words, during the promotional period 3.7 Flash costs half of what 3.6 Flash costs while being stronger across the board — the next few months are the best time to migrate.

Stack the top-up promotions

APIYI always matches official pricing, and discounts come from top-up bonuses, which stack on top of the promotional rate above: 📖 Top-up promotion details

Summary

The Flash tier has improved steadily across three consecutive versions, and 3.7 is the largest step yet — coding benchmarks up by 10+ points across the board, agents and business automation close to doubled, and a unit price that is currently half the previous generation. The recommendation is simple: make gemini-3.7-flash your default for Gemini text workloads. The only things worth validating separately are chart-heavy analysis (CharXiv regressed slightly) and whichever built-in tool you depend on most.

Sources: Google DeepMind official model card deepmind.google/models/model-cards/gemini-3-7-flash/ and the Gemini API documentation ai.google.dev/gemini-api/docs/latest-model. Released August 13, 2026; data collected August 14, 2026. Pricing follows the live APIYI model pricing page.

Changelog

AI Radar

Gemini 3.7 Flash Is Live: Big Coding and Agent Gains

Key Takeaways

Background

Deep Dive

Official benchmarks (3.7 Flash vs 3.6 Flash)

Technical specs

In Practice

Recommended use cases

Code examples

Best practices

Pricing & Availability

Stack the top-up promotions

Summary

​Key Takeaways

​Background

​Deep Dive

​Official benchmarks (3.7 Flash vs 3.6 Flash)

​Technical specs

​In Practice

​Recommended use cases

​Code examples

​Best practices

​Pricing & Availability

​Stack the top-up promotions

​Summary

Key Takeaways

Background

Deep Dive

Official benchmarks (3.7 Flash vs 3.6 Flash)

Technical specs

In Practice

Recommended use cases

Code examples

Best practices

Pricing & Availability

Stack the top-up promotions

Summary