Model Drop

Gemini 2.5 Flash

April 17, 2025 2 min read

Gemini 2.5 Flash, which Google DeepMind shipped as a public preview on April 17, 2025 (generally available June 17, 2025), is its first fully hybrid reasoning model: a cheap, fast Flash model that can turn thinking on or off and cap it with a budget.

Model: Gemini 2.5 Flash (gemini-2.5-flash-preview-04-17)
Maker: Google DeepMind
Ship date: April 17, 2025
Pricing: $0.15 / $0.60 per M tokens ($3.50 output with thinking on)
Context: 1M tokens
Weights: Closed
Where you can use it:
- Gemini app, for everyone
- Google AI Studio and Vertex AI (preview)
When to reach for it:
- Cheap reasoning at high volume
- Drop-in upgrade from Gemini 2.0 Flash
- Skip it for the hardest math and coding
How it compares:
- Cheaper than o4-mini per token
- Trails o4-mini on math and code
What makes it different:
- Thinking can be turned on or off
- Thinking budget caps reasoning tokens
Early reactions: Early testers called the split pricing for thinking and non-thinking output a little complicated, and saw it still think briefly with the budget at zero.
Read more: Start building with Gemini 2.5 Flash, Gemini 2.5 Flash is now in preview

The benchmarks

Gemini 2.5 Flash Preview (04-17) with thinking at default settings, single attempt; Google picked Gemini 2.0 Flash (non-thinking), o4-mini, Claude 3.7 Sonnet (extended thinking: 64K on GPQA and MMMU, 16K on HLE), Grok 3 Beta (extended thinking, except SimpleQA) and DeepSeek R1, all from the makers' own reports or public leaderboards.

Google DeepMind's numbers

Humanity's Last Exam (no tools) (expert questions): 12.1%. −2.2 under o4-mini (14.3%), +3.2 over Claude 3.7 Sonnet (8.9%), +7.0 over Gemini 2.0 Flash (5.1%).
GPQA Diamond (PhD-level science): 78.3%. −3.1 under o4-mini (81.4%), +0.1 over Claude 3.7 Sonnet (78.2%), −1.9 under Grok 3 Beta (80.2%), +6.8 over DeepSeek R1 (71.5%).
AIME 2025 (competition math): 78.0%. −14.7 under o4-mini (92.7%), +0.7 over Grok 3 Beta (77.3%), +8.0 over DeepSeek R1 (70.0%), +50.5 over Gemini 2.0 Flash (27.5%).
LiveCodeBench (v5, coding problems): 63.5%. −7.1 under Grok 3 Beta (70.6%), −0.8 under DeepSeek R1 (64.3%), +29.0 over Gemini 2.0 Flash (34.5%).
Aider Polyglot (code editing, whole format): 51.1%. −17.8 under o4-mini (68.9%), +28.9 over Gemini 2.0 Flash (22.2%).
MMMU (visual reasoning): 76.7%. −4.9 under o4-mini (81.6%), +1.7 over Claude 3.7 Sonnet (75.0%), +0.7 over Grok 3 Beta (76.0%), +5.0 over Gemini 2.0 Flash (71.7%).
SimpleQA (factual recall): 29.7%. −13.9 under Grok 3 Beta (43.6%), −0.4 under DeepSeek R1 (30.1%), −0.2 under Gemini 2.0 Flash (29.9%).
MRCR (1M) (long-context recall, pointwise): 66.3%. +18.1 over Gemini 2.0 Flash (48.2%).

Independent numbers

Artificial Analysis Intelligence Index (current v4.3.2, AA estimate): 11.7. −5.0 under o4-mini (high) (16.7), −6.0 under Claude 3.7 Sonnet (thinking) (17.7), +0.3 over DeepSeek R1 (11.4), +2.8 over Gemini 2.0 Flash (8.9).