Introduction
This entry covers Gemini 2.5 as announced at I/O 2025. Google has since released newer generations, including Gemini 3, Gemini 3.1 Pro, and Gemini 3.5; see Google's official Gemini models documentation for the latest models.
Google unveiled the Gemini 2.5 series during the I/O 2025 keynote. Building on the March release of 2.5 Pro, the update improves reasoning, multimodal comprehension, and developer tooling. These enhancements now power Search's AI Mode, the Gemini app, and enterprise services such as Vertex AI.
Key Performance Enhancements
- Deep Think — an experimental reasoning mode that evaluates multiple hypotheses before answering. It notably boosts math and coding performance.
- Native audio output delivers expressive, emotion-aware speech for more natural conversations.
- Enhanced security mitigates prompt injection attacks when using tools.
- Agent Mode via Project Mariner enables autonomous web browsing and transactions.
Together, these changes make Gemini interactions more reliable and capable.
Model Variants
Gemini 2.5 ships in three main variants:
- 2.5 Pro — optimized for advanced coding and reasoning with Deep Think and a 1 million token context window.
- 2.5 Flash — tuned for speed and lower cost, using 20–30 % fewer tokens while improving reasoning scores.
- 2.5 Flash Lite — lightweight version for mobile and embedded deployments, prioritizing minimal latency and cost.
Additional current family context: The Gemini 2.5 family ultimately included these three variants. Flash-Lite became the family's lowest-cost, lowest-latency option for high-volume tasks such as translation and classification.
The dates below distinguish each model's initial preview from the updated versions shown at I/O. Google introduced 2.5 Pro Experimental on March 25, 2025, released the first 2.5 Flash preview on April 17, 2025, and presented updated Pro and Flash previews at I/O on May 20. Both models became generally available on June 17, when Google also introduced Flash-Lite in preview.
2.5 Pro Model Card
| Attribute | Details |
|---|---|
| Developer | Google DeepMind |
| Release Date | May 20, 2025 |
| Access | Public preview on Vertex AI and Google AI Studio; general availability shortly after June 2025 |
| Context Window | 1 000 000 tokens with 2 000 000 planned in a later update |
| Input Modalities | Text, code, images, audio, and video |
| Key Features | Experimental Deep Think mode, native emotion-aware audio, multilingual support, enhanced security against indirect prompt injections |
| Benchmarks | Leading results on LiveCodeBench, USAMO, and MMMU |
| Pricing | $2.50 per 1 000 000 input tokens; $15.00 per 1 000 000 output tokens |
| Limitations | Deep Think limited to trusted testers; occasional hallucinations |
Current 2.5 Pro timeline and pricing context
Additive corrections that do not replace the I/O 2025 launch card above.
| Attribute | Current context |
|---|---|
| Release Timeline | Initial experimental release March 25, 2025; updated I/O preview May 20; general availability June 17, 2025 |
| Access | Initially available in Google AI Studio and the Gemini app; later previewed on Vertex AI before general availability |
| Context Window | 1 000 000 tokens with 2 000 000 planned in a later update |
| Input Modalities | Text, code, images, audio, and video |
| Key Features | Experimental Deep Think mode, native emotion-aware audio, multilingual support, enhanced security against indirect prompt injections |
| Benchmarks | Google-reported I/O 2025 results on LiveCodeBench, USAMO, and MMMU |
| Pricing | Per 1 000 000 tokens: prompts up to 200K cost $1.25 input and $10 output; prompts over 200K cost $2.50 input and $15 output |
| Limitations | Deep Think limited to trusted testers; occasional hallucinations |
2.5 Flash Model Card
| Attribute | Details |
|---|---|
| Developer | Google DeepMind |
| Release Date | May 20, 2025 |
| Access | Preview in Google AI Studio, Vertex AI, and the Gemini app; general availability early June 2025 |
| Context Window | Inherits Gemini's long-context architecture (token limit undisclosed) |
| Input Modalities | Text, images, code, audio, and video |
| Key Features | Optimized for low latency and cost, using 20–30 % fewer tokens; configurable thinking budget per request |
| Benchmarks | Competitive reasoning, multimodal understanding, and code generation with sub-200 ms responses |
| Pricing | Consumption-based pricing with lower cost per token than Pro |
| Limitations | Slightly lower accuracy than Pro on complex tasks |
Current 2.5 Flash timeline and context
| Attribute | Current context |
|---|---|
| Release Timeline | Initial preview April 17, 2025; updated I/O preview May 20; general availability June 17, 2025 |
| Access | Previewed in Google AI Studio, Vertex AI, and the Gemini app before general availability |
| Context Window | 1 000 000 tokens |
| Input Modalities | Text, images, code, audio, and video |
| Key Features | Optimized for low latency and cost, using 20–30 % fewer tokens; configurable thinking budget per request |
| Benchmarks | Google-reported I/O 2025 improvements in reasoning, multimodal understanding, coding, and long-context performance |
| Pricing | Consumption-based pricing with lower cost per token than Pro |
| Limitations | Slightly lower accuracy than Pro on complex tasks |
2.5 Flash Lite Model Card
| Attribute | Details |
|---|---|
| Developer | Google DeepMind |
| Release Date | May 20, 2025 |
| Access | Limited preview on Android via the Gemini app and Google AI Studio |
| Context Window | 128K tokens |
| Input Modalities | Text and images |
| Key Features | On-device optimized model for minimal latency and cost |
| Benchmarks | Similar accuracy to Flash on straightforward tasks |
| Pricing | Free for consumer mobile use; API pricing TBA |
| Limitations | Reduced accuracy on complex tasks and smaller context |
Current 2.5 Flash-Lite correction context
Google introduced Gemini 2.5 Flash-Lite on June 17, 2025, after the May 20 I/O update covered by this entry.
| Attribute | Current context |
|---|---|
| Developer | Google DeepMind |
| Release Date | June 17, 2025 |
| Access | Preview in Google AI Studio and Vertex AI; custom versions also used in Google Search |
| Context Window | 1 000 000 tokens |
| Input Modalities | Multimodal input |
| Key Features | Adjustable thinking budgets, Google Search and code-execution tools, and optimization for high-volume, latency-sensitive tasks |
| Benchmarks | Google reported higher quality than 2.0 Flash-Lite across coding, math, science, reasoning, and multimodal evaluations, plus lower latency |
| Pricing | Introduced as the most cost-efficient model in the Gemini 2.5 family |
| Limitations | Preview release at introduction |
Benchmarks and Learning Focus
Gemini 2.5 Pro leads major leaderboards like WebDev Arena and LMArena, demonstrating strong coding and reasoning. Deep Think sets records on the USAMO exam, LiveCodeBench, and MMMU. Using the LearnLM approach, the model excels on pedagogical benchmarks, outperforming peers across five learning science principles.
Additional dated benchmark context: In its May 20, 2025 I/O announcement, Google reported that Gemini 2.5 Pro led WebDev Arena and all LMArena leaderboards at that launch snapshot. Google also reported that the experimental Deep Think mode posted strong results on the 2025 USAMO exam, LiveCodeBench, and MMMU. Using the LearnLM approach, Google reported that the model outperformed peers across five learning-science principles. These provider-reported launch results should not be read as current leaderboard records.
Developer and Enterprise Experience
Gemini API and Vertex AI SDK now provide thought summaries showing each model's reasoning and tool usage. Enterprises can audit easier and set Thinking Budgets up to 32K tokens. MCP tools offer extensibility, and early testers on Vertex AI can evaluate Deep Think in a secure environment.
Access Plans
The AI Premium tier has been renamed Google AI Pro, adding Flow (Veo 2) and early Chrome integration. The Gemini AI Ultra subscription grants VIP access to Deep Think, Veo 3 video generation, Project Mariner, and more.
Gemini AI Ultra Subscription Card
| Attribute | Details |
|---|---|
| Plan Name | Gemini AI Ultra |
| Launch Date | May 20, 2025 |
| Price | US$249.99 per month with a 50 % introductory discount in the U.S. |
| Access | VIP access to Gemini 2.5 Pro Deep Think, Veo 3, Flow, NotebookLM, and Project Mariner |
| Storage | 30 TB across Google Photos, Drive, and Gmail |
| Extras | YouTube Premium subscription included |
| Limitations | Paid subscription required; initial roll-out limited to the U.S. |
Current Google AI Ultra tier context
Additive naming and pricing corrections.
| Attribute | Current context |
|---|---|
| Plan Name | Google AI Ultra |
| Launch Date | May 20, 2025 |
| Price | US$249.99 per month at launch (May 2025), with a 50 % introductory discount in the U.S.; at I/O on May 19, 2026, Google lowered the price to US$200 per month while keeping the same capabilities and 20x Pro usage |
| Access | VIP access to Gemini 2.5 Pro Deep Think, Veo 3, Flow, NotebookLM, and Project Mariner |
| Storage | 30 TB across Google Photos, Drive, and Gmail |
| Extras | YouTube Premium subscription included |
| Limitations | Paid subscription required; initial roll-out limited to the U.S. |
At the same I/O 2026 event, Google introduced a separate, lower-cost AI Ultra tier at US$100 per month, offering 5x Pro usage and 20 TB of storage — distinct from the top-tier plan's 20x Pro usage and 30 TB. See Google's official pricing announcement for details.
Availability and Pricing
2.5 Flash is live in preview via the Gemini app, Google AI Studio, and Vertex AI. General availability arrives in early June with 2.5 Pro shortly after. Gemini AI Ultra launches in the US with a 50 % discount and will expand to 70+ countries. Students in select markets receive a free year of AI Pro.
Additional current availability context: Gemini 2.5 Flash and Pro became generally available in Google AI Studio and Vertex AI on June 17, 2025. Flash-Lite entered preview on those developer platforms the same day, while Flash and Pro were also available in the Gemini app. At its May 2025 launch, Google AI Ultra was offered in the U.S. with a 50 % introductory discount, and Google said it would expand to more than 70 countries. Students in select markets were offered a free year of AI Pro.
Implications and Next Steps
With Gemini 2.5 integrated into Search's AI Mode, Chrome, and Google Workspace, users gain more proactive assistance. Upcoming features include deep search, AI-generated charts, and agentic checkout. Enterprises are testing extraction agents in Box and workflows in Geotab, aiming for over 90 % document accuracy. Google plans wider Deep Think access after further safety evaluations.