Model Drop

Gemini 3 Flash

December 17, 2025 3 min read

Gemini 3 Flash, which Google DeepMind shipped on December 17, 2025, is the fast, low-cost member of the Gemini 3 family, bringing near-Pro reasoning at Flash speed and becoming the default model in the Gemini app.

Model: Gemini 3 Flash (gemini-3-flash-preview)
Maker: Google DeepMind
Ship date: December 17, 2025
Pricing: $0.50 / $3 per M tokens
Context: 1M tokens
Weights: Closed
Where you can use it:
- Gemini app and AI Mode in Search
- Google AI Studio, Gemini CLI, Antigravity
- Vertex AI and Gemini Enterprise (preview)
When to reach for it:
- Fast agentic coding loops
- Video, image and document analysis
- Skip it for the hardest abstract reasoning
How it compares:
- Near Gemini 3 Pro at a quarter the price
- Pricier than Gemini 2.5 Flash per token
What makes it different:
- Four thinking levels, minimal to high
- Uses fewer tokens than Gemini 2.5 Pro
Early reactions: Early testers praised its coding speed and price, though some noted it costs more than 2.5 Flash and drops image segmentation.
Read more: Gemini 3 Flash: frontier intelligence built for speed, Gemini 3 Flash model card

The benchmarks

Gemini 3 Flash Preview at its default high thinking level and default API settings, single attempt; Google picked Gemini 3 Pro, Gemini 2.5 Flash and 2.5 Pro (all thinking), Claude Sonnet 4.5 (thinking, high where reported), GPT-5.2 at xhigh and Grok 4.1 Fast (reasoning, using third-party evals since xAI hadn't published them), mostly self-reported, with some rival scores run by Google. Terminal-Bench 2.0 scores come from the public leaderboard with the Terminus-2 harness, and Gemini 3 Flash's own submission was still pending. The LMArena row is the first snapshot listing it (dated December 16, 2025, the day before launch, on 3,824 votes; it still read 1470 on December 18), and the ARC-AGI-1 row uses ARC Prize's high thinking run.

Google DeepMind's numbers

Humanity's Last Exam (no tools) (expert questions): 33.7%. −3.8 under Gemini 3 Pro (37.5%), +20.0 over Claude Sonnet 4.5 (13.7%), −0.8 under GPT-5.2 (34.5%), +16.1 over Grok 4.1 Fast (17.6%).
GPQA Diamond (PhD-level science): 90.4%. −1.5 under Gemini 3 Pro (91.9%), +7.0 over Claude Sonnet 4.5 (83.4%), −2.0 under GPT-5.2 (92.4%), +6.1 over Grok 4.1 Fast (84.3%).
ARC-AGI-2 (visual reasoning puzzles): 33.6%. +2.5 over Gemini 3 Pro (31.1%), +20.0 over Claude Sonnet 4.5 (13.6%), −19.3 under GPT-5.2 (52.9%).
MMMU-Pro (multimodal reasoning): 81.2%. +0.2 over Gemini 3 Pro (81.0%), +13.2 over Claude Sonnet 4.5 (68.0%), +1.7 over GPT-5.2 (79.5%), +18.2 over Grok 4.1 Fast (63.0%).
SWE-bench Verified (real GitHub fixes): 78.0%. +1.8 over Gemini 3 Pro (76.2%), +0.8 over Claude Sonnet 4.5 (77.2%), −2.0 under GPT-5.2 (80.0%), +27.4 over Grok 4.1 Fast (50.6%), +17.6 over Gemini 2.5 Flash (60.4%).
Terminal-Bench 2.0 (agentic terminal coding): 47.6%. −6.6 under Gemini 3 Pro (54.2%), +4.8 over Claude Sonnet 4.5 (42.8%), +30.7 over Gemini 2.5 Flash (16.9%).
MCP Atlas (multi-step MCP workflows): 57.4%. +3.3 over Gemini 3 Pro (54.1%), +13.6 over Claude Sonnet 4.5 (43.8%), −3.2 under GPT-5.2 (60.6%).
SimpleQA Verified (factual recall): 68.7%. −3.4 under Gemini 3 Pro (72.1%), +39.4 over Claude Sonnet 4.5 (29.3%), +30.7 over GPT-5.2 (38.0%), +49.2 over Grok 4.1 Fast (19.5%).

Independent numbers

Artificial Analysis Intelligence Index (current v4.3.2, AA estimate): 26.3. −4.1 under GPT-5.2 (xhigh) (30.4), +5.6 over Claude Sonnet 4.5 (thinking) (20.7), +5.9 over Grok 4.1 Fast (reasoning) (20.4), +9.5 over GPT-5 mini (high) (16.8).
LMArena text Elo (human preference votes): 1470. +51 over GPT-5.2 (high) (1419), +43 over Claude Sonnet 4.5 (thinking) (1427), +60 over Grok 4.1 Fast (reasoning) (1410), +86 over Claude Haiku 4.5 (1384).
ARC-AGI-1 (ARC Prize, semi-private set): 84.7%. +9.7 over Gemini 3 Pro (75.0%), −1.5 under GPT-5.2 (xhigh) (86.2%), +21.0 over Claude Sonnet 4.5 (thinking) (63.7%).