Model Drop

Gemini 3.6 Flash

July 21, 2026 2 min read

Gemini 3.6 Flash, which Google DeepMind shipped on July 21, 2026, is a cheaper, more token-efficient update to its Flash workhorse model for agents, coding and knowledge work.

Model: Gemini 3.6 Flash (gemini-3.6-flash)
Maker: Google DeepMind
Ship date: July 21, 2026
Pricing: $1.50 / $7.50 per M tokens
Context: 1M tokens
Where you can use it:
- Gemini app and the Gemini Enterprise app
- Google AI Studio and Android Studio
- Google Antigravity and Gemini Enterprise
When to reach for it:
- Computer-use agents that click through UIs
- High-volume agent pipelines on a budget
- Skip it for long, hard software engineering
How it compares:
- Cheaper output than Claude Sonnet 5
- Trails GPT-5.6 Luna, Grok 4.5 on coding
What makes it different:
- Uses fewer output tokens than 3.5 Flash
- Computer use is a built-in API tool
Early reactions: Hacker News users loved its SVG drawing, but many said rivals beat it on cost for performance at this price.
Read more: Introducing Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber, Gemini 3.6 Flash model card

The benchmarks

Gemini 3.6 Flash at default API settings (medium thinking by default; high thinking on DeepSWE); Google picked GPT-5.6 Luna, Grok 4.5 and Claude Sonnet 5 at max reasoning, mostly self-reported or from public leaderboards, with some rival scores (OSWorld for Luna, CharXiv) run by Google.

Google DeepMind's numbers

SWE-bench Pro (harder agentic coding): 58.7%. −4.0 under GPT-5.6 Luna (62.7%), −6.0 under Grok 4.5 (64.7%), −4.5 under Claude Sonnet 5 (63.2%), +3.6 over Gemini 3.5 Flash (55.1%).
DeepSWE v1.1 (long-horizon software engineering): 49%. −18 under GPT-5.6 Luna (67%), −5 under Grok 4.5 (54%), −5 under Claude Sonnet 5 (54%), +12 over Gemini 3.5 Flash (37%).
Terminal-Bench 2.1 (agentic terminal coding): 78.0%. −6.7 under GPT-5.6 Luna (84.7%), −5.3 under Grok 4.5 (83.3%), −2.4 under Claude Sonnet 5 (80.4%), +1.8 over Gemini 3.5 Flash (76.2%).
MLE-Bench (machine learning engineering): 63.9%. +16.3 over GPT-5.6 Luna (47.6%), +20.7 over Grok 4.5 (43.2%), −3.0 under Claude Sonnet 5 (66.9%).
GDPval-AA v2 (knowledge work Elo): 1421. −163 under GPT-5.6 Luna (1584), −114 under Grok 4.5 (1535), −186 under Claude Sonnet 5 (1607), +72 over Gemini 3.5 Flash (1349).
OSWorld-Verified (computer use): 83.0%. +10.4 over GPT-5.6 Luna (72.6%), +1.8 over Claude Sonnet 5 (81.2%), +4.6 over Gemini 3.5 Flash (78.4%).
CharXiv Reasoning (chart reading, no tools): 85.2%. +2.5 over GPT-5.6 Luna (82.7%), +3.6 over Grok 4.5 (81.6%), +8.2 over Claude Sonnet 5 (77.0%).
GDM-MRCR v2 (long-context recall, 128K): 91.8%. +17.0 over GPT-5.6 Luna (74.8%), +10.4 over Grok 4.5 (81.4%), +20.2 over Claude Sonnet 5 (71.6%).

Independent numbers

Artificial Analysis Intelligence Index (current v4.3.2): 34.0. −3.3 under GPT-5.6 Luna (max) (37.3), −4.8 under Grok 4.5 (high) (38.8), −4.2 under Claude Sonnet 5 (max) (38.2).
ARC-AGI-2 (ARC Prize verified puzzles): 60.4%. +0.9 over GPT-5.6 Luna (max) (59.5%), +7.8 over Grok 4.5 (high) (52.6%).