Model Drop

Gemini 3.7 Flash

August 13, 2026 3 min read

Gemini 3.7 Flash, which Google DeepMind shipped on August 13, 2026, is a faster-follow update to its Flash workhorse model for coding and agents, launched three weeks after 3.6 Flash at an introductory price of half 3.6 Flash's original price.

Model: Gemini 3.7 Flash (gemini-3.7-flash)
Maker: Google DeepMind
Ship date: August 13, 2026
Pricing: $0.75 / $3.75 per M tokens (introductory through December 31, 2026; $1.50 / $7.50 from January 1, 2027)
Context: 1M tokens
Weights: Closed
Where you can use it:
- Gemini app via Spark, for AI Pro and Ultra
- Google AI Studio, Gemini API, Android Studio
- Google Antigravity and Gemini Enterprise
When to reach for it:
- Cheap, high-volume coding agents
- Document-heavy finance and legal work
- Skip it for the hardest terminal tasks
How it compares:
- Far cheaper per token than Claude Sonnet 5
- Trails GPT-5.6 Terra on hard coding
What makes it different:
- Intro price is half 3.6 Flash's launch price
- Shipped just three weeks after 3.6 Flash
Early reactions: Developers cheered the low price and strong security triage results, while some testers called it disappointing next to GPT-5.6 Terra.
Read more: Introducing Gemini 3.7 Flash, Gemini 3.7 Flash model card

The benchmarks

Gemini 3.7 Flash at default Gemini API settings (high thinking on DeepSWE, where 3.6 Flash's 48.6% is from the table and model card, also at high; the post's prose says 49.0%); Google picked Claude Sonnet 5, GPT-5.6 Terra and Muse Spark 1.2 at their maximum reported reasoning settings, mostly self-reported or from public leaderboards, but Google ran some rival scores itself (GDP.pdf for Sonnet 5, CharXiv for Terra, HLE-Verified for all models, where content filters blocked a significant share of Sonnet 5's questions), and the Gemini Terminal-Bench 3.0 runs used Google's own harness with two tasks modified. The independent rows use Gemini 3.7 Flash at high.

Google DeepMind's numbers

DeepSWE v1.1 (long-horizon software engineering): 65.3%. +16.7 over Gemini 3.6 Flash (48.6%), +11.5 over Claude Sonnet 5 (53.8%), −4.3 under GPT-5.6 Terra (69.6%), +10.4 over Muse Spark 1.2 (54.9%).
Terminal-Bench 2.1 (agentic terminal coding): 85.8%. +7.8 over Gemini 3.6 Flash (78.0%), +5.4 over Claude Sonnet 5 (80.4%), −1.6 under GPT-5.6 Terra (87.4%), +2.9 over Muse Spark 1.2 (82.9%).
Terminal-Bench 3.0 (general agent tasks): 14.9%. +9.5 over Gemini 3.6 Flash (5.4%), +0.3 over Claude Sonnet 5 (14.6%), −5.9 under GPT-5.6 Terra (20.8%).
AutomationBench (business workflow automation): 30.4%. +13.4 over Gemini 3.6 Flash (17.0%), +19.7 over Claude Sonnet 5 (10.7%), +6.8 over GPT-5.6 Terra (23.6%).
GDPval-AA v2 (knowledge work Elo): 1525. +103 over Gemini 3.6 Flash (1422), −73 under Claude Sonnet 5 (1598), −53 under GPT-5.6 Terra (1578), −103 under Muse Spark 1.2 (1628).
GDP.pdf (expert PDF comprehension): 34.0%. +12.0 over Gemini 3.6 Flash (22.0%), +6.0 over Claude Sonnet 5 (28.0%), +9.3 over GPT-5.6 Terra (24.7%), +18.0 over Muse Spark 1.2 (16.0%).
HLE-Verified (expert reasoning questions): 53.6%. +2.4 over Gemini 3.6 Flash (51.2%), +22.6 over Claude Sonnet 5 (31.0%), +2.5 over GPT-5.6 Terra (51.1%).
CharXiv Reasoning (chart reading, no tools): 84.5%. −0.7 under Gemini 3.6 Flash (85.2%), +7.5 over Claude Sonnet 5 (77.0%), −1.4 under GPT-5.6 Terra (85.9%).

Independent numbers

Artificial Analysis Intelligence Index (current v4.3.2): 39.1. −3.0 under GPT-5.6 Terra (max) (42.1), +0.9 over Claude Sonnet 5 (max) (38.2), −0.5 under Muse Spark 1.2 (xhigh) (39.6).
ARC-AGI-2 (ARC Prize verified puzzles): 84.6%. +0.7 over GPT-5.6 Terra (max) (83.9%), +24.2 over Gemini 3.6 Flash (high) (60.4%).
ARC-AGI-1 (ARC Prize verified puzzles): 95.5%. −1.0 under GPT-5.6 Terra (max) (96.5%), +4.3 over Gemini 3.6 Flash (high) (91.2%).