Last week, Google released Gemini Embedding 2, the first natively multimodal embedding model built on the Gemini architecture. If you work with embeddings in any capacity, this deserves your attention. It has the potential to significantly disrupt the multi-model embedding pipelines that most teams rely on today.

Until now, the flagship embedding models from OpenAI, Cohere, and Voyage were primarily text-based. A few multimodal options existed — CLIP for image-text alignment, Voyage Multimodal 3.5 for images and video — but none covered the full spectrum of modalities in a single, unified vector space. Audio typically had to be transcribed before embedding. Video required frame extraction combined with separate transcript embeddings. Images lived in their own vector space entirely.

Gemini Embedding 2 changes that equation. One model, one API call, one vector space.

Let’s dig into what’s new.

What Is Gemini Embedding 2?

Gemini Embedding 2 (gemini-embedding-2-preview) is Google DeepMind’s first fully multimodal embedding model. It takes text, images, video clips, audio recordings, and PDF documents and converts all of them into vectors that live in the same shared semantic space.

Unlike earlier multimodal approaches such as CLIP, which pair a vision encoder with a text encoder and align them with contrastive learning at the end, Gemini Embedding 2 is built on the Gemini foundation model itself. This means it inherits deep cross-modal understanding from the ground up.

Image generated using Nano Banana

Practical example: Imagine you’re building a Learning Management System (LMS) with video tutorials, audio lectures, and written guides. With Gemini Embedding 2, you can store embeddings for all of this content in a single vector space and build a RAG-based chatbot that retrieves relevant chunks from videos, audio, and documents alike. Previously, this required a multi-layered embedding pipeline — and even then, it only captured transcripts, missing the visual context of a video or the tone of a speaker’s voice.

The model uses Matryoshka Representation Learning, which means you don’t have to use all 3072 dimensions if you don’t need them. You can scale down to 1536 or 768 and still get usable results.

Matryoshka Representation Learning (MRL) is a technique for training embedding models so that the learned representations are useful not only at their full dimensionality but also at various smaller dimensions — nested inside one another like Russian matryoshka dolls. During training, the loss function is computed not just on the full embedding but also on multiple prefixes of the embedding vector. This encourages the model to pack the most important information into the earliest dimensions, with each subsequent dimension adding finer-grained detail — a coarse-to-fine structure.

Supported Modalities & Input Limits

The model accepts five types of input, all mapped into the same embedding space:

How It Compares to Existing Models

TLDR: Google’s new Gemini Embedding 2 model tops its competitors (its own predecessor, Amazon Nova 2, and Voyage Multimodal 3.5) across nearly every modality: text, image, video, and speech. It leads most convincingly in video retrieval and image-text matching. The only benchmark where it doesn’t win is document retrieval, where Voyage edges slightly ahead. Speech-text retrieval is a category Gemini owns alone since no competitor even supports it.

Google published benchmark comparisons against its own legacy models, Amazon Nova 2 Multimodal Embeddings, and Voyage Multimodal 3.5. Here’s the full picture:

Text-Text

Gemini Embedding 2 leads on multilingual text by a comfortable margin and jumps 8 points over its own predecessor on code retrieval. Neither Amazon Nova 2 nor Voyage report code scores.

Text-Image

A clear lead in text-to-image retrieval — over 9 points ahead of the nearest competitor on both benchmarks.

Image-Text

Image-to-text retrieval shows the widest gaps — nearly 15 points ahead of Amazon Nova 2 on Docci.

Text-Document

The one benchmark where Voyage Multimodal 3.5 edges ahead (self-reported). Document retrieval is close between the top models.

Text-Video

Video retrieval is where Gemini Embedding 2 pulls furthest ahead — over 17 points above Voyage on Youcook2 and over 13 points on Vatex.

Speech-Text

Speech-text retrieval is entirely uncontested — neither Amazon nor Voyage support it. This is a category Gemini Embedding 2 owns outright.

– score not available ** self-reported *** voyage-3.5 **** ASR model converts audio queries to text

Pricing

The model is currently free during public preview. Once on the paid tier, here’s the breakdown:

Getting Started

The model is available now in public preview via the Gemini API and Vertex AI under the model ID gemini-embedding-2-preview. It integrates with LangChain, LlamaIndex, Haystack, Weaviate, Qdrant, ChromaDB, and Vector Search.

from google import genai
from google.genai import types

# For Vertex AI:
# PROJECT_ID='<add_here>'
# client = genai.Client(vertexai=True, project=PROJECT_ID, location='us-central1')

client = genai.Client()

with open("example.png", "rb") as f:
    image_bytes = f.read()

with open("sample.mp3", "rb") as f:
    audio_bytes = f.read()

# Embed text, image, and audio 
result = client.models.embed_content(
    model="gemini-embedding-2-preview",
    contents=[
        "What is the meaning of life?",
        types.Part.from_bytes(
            data=image_bytes,
            mime_type="image/png",
        ),
        types.Part.from_bytes(
            data=audio_bytes,
            mime_type="audio/mpeg",
        ),
    ],
)

print(result.embeddings)

Try it out here!

We’ve built a demo app where you can test out the multimodal retrieval performance of gemini-embedding-2.

You can get the API Key by logging into aistudio.google.com.

Limitations to Watch

• The model is still in public preview (the “preview” tag means pricing and behavior may change before GA).
• Video input is capped at 120 seconds and audio at 80 seconds.
• Performance on niche domains like financial QA is weaker; evaluate against your specific data before committing.
• For pure text pipelines with no multimodal plans, the cost premium over text-only models may not be justified.

The Bottom Line

Gemini Embedding 2 isn’t just an incremental improvement, it’s a category shift. For teams building multimodal RAG systems, semantic search across media types, or unified knowledge bases, it collapses what used to be a multi-model, multi-pipeline problem into a single API call. If your data spans more than just text, this is the model to evaluate first.

Building multimodal RAG shouldn’t mean stitching together embedding models, vector databases, and retrieval logic from scratch. If you want a managed RAG-as-a-Service solution that handles the embedding pipeline for you, sign up for the free trial at Cody and start building today.

The post Gemini Embedding 2: Google’s First Multimodal Embedding Model appeared first on Cody - The AI Trained on Your Business.

Not wanting to fall behind, OpenAI launched GPT-4.5, their largest and most advanced chat model so far. This model is great at recognizing patterns, making connections, and coming up with creative ideas. Early tests show that interacting with GPT-4.5 feels very natural, thanks to its wide range of knowledge and improved understanding of what users mean. OpenAI emphasizes GPT-4.5’s significant improvements in learning without direct supervision, designed for smooth collaboration with humans.

These AI systems are not just impressive technology; they are changing how businesses operate, speeding up scientific discoveries, and transforming creative projects. As AI becomes a normal part of daily life, models like Gemini 2.5 Pro and GPT-4.5 are expanding what we think is possible. With better reasoning skills, less chance of spreading false information, and mastery over complex problems, they are paving the way for AI systems that truly support human progress.

Understanding Gemini 2.5 Pro

On March 25, 2025, Google officially unveiled Gemini 2.5 Pro, described as their “most intelligent AI model” to date. This release marked a significant milestone in Google’s AI development journey, coming after several iterations of their 2.0 models. The release strategy began with the experimental version first, giving Gemini Advanced subscribers early access to test its capabilities.

What separates Gemini 2.5 Pro from its predecessors is its fundamental architecture as a “thinking model.” Unlike previous generations that primarily relied on trained data patterns, this model can actively reason through its thoughts before responding, mimicking human problem-solving processes. This represents a significant advancement in how AI systems process information and generate responses.

Key Features and Capabilities:

1. Enhanced reasoning abilities – Capable of step-by-step problem solving across complex domains
2. Expanded context window – 1 million token capacity (with plans to expand to 2 million)
3. Native multimodality – Seamlessly processes text, images, audio, video, and code
4. Advanced code capabilities – Significant improvements in web app creation and code transformation

Gemini 2.5 Pro has established itself as a performance leader, debuting at the #1 position on the LMArena leaderboard. It particularly excels in benchmarks requiring advanced reasoning, scoring an industry-leading 18.8% on Humanity’s Last Exam without using external tools. In mathematics and science, it demonstrates remarkable competence with scores of 86.7% on AIME 2025 and 79.7% on GPQA diamond respectively.

Compared to previous Gemini models, version 2.5 Pro represents a substantial leap forward. While Gemini 2.0 introduced important foundational capabilities, 2.5 Pro combines a significantly enhanced base model with improved post-training techniques. The most notable improvements appear in coding performance, reasoning depth, and contextual understanding—areas where earlier versions showed limitations.

Exploring GPT-4.5

In April 2025, OpenAI introduced GPT-4.5, describing it as their “largest and most advanced chat model to date,” signifying a noteworthy achievement in the evolution of large language models. This research preview sparked immediate excitement within the AI community, with initial tests indicating that interactions with the model feel exceptionally natural, thanks to its extensive knowledge base and enhanced ability to comprehend user intent.

GPT-4.5 showcases significant advancements in unsupervised learning capabilities. OpenAI realized this progress by scaling both computational power and data inputs, alongside employing innovative architectural and optimization strategies. The model was trained on Microsoft Azure AI supercomputers, continuing a partnership that has enabled OpenAI to push the boundaries of possibility.

Core Improvements and Capabilities:

1. Enhanced pattern recognition – Significantly improved ability to recognize patterns, draw connections, and generate creative insights
2. Reduced hallucinations – Less likely to generate false information compared to previous models like GPT-4o and o1
3. Improved “EQ” – Greater emotional intelligence and understanding of nuanced human interactions
4. Advanced steerability – Better understanding of and adherence to complex user instructions

OpenAI has placed particular emphasis on training GPT-4.5 for human collaboration. New techniques enhance the model’s steerability, understanding of nuance, and natural conversation flow. This makes it particularly effective in writing and design assistance, where it demonstrates stronger aesthetic intuition and creativity than previous iterations.

In real-world applications, GPT-4.5 shows remarkable versatility. Its expanded knowledge base and improved reasoning capabilities make it suitable for a wide range of tasks, from detailed content creation to sophisticated problem-solving. OpenAI CEO Sam Altman has described the model in positive terms, highlighting its “unique effectiveness” despite not leading in all benchmark categories.

The deployment strategy for GPT-4.5 reflects OpenAI’s measured approach to releasing powerful AI systems. Initially available to ChatGPT Pro subscribers and developers on paid tiers through various APIs, the company plans to gradually expand access to ChatGPT Plus, Team, Edu, and Enterprise subscribers. This phased rollout allows OpenAI to monitor performance and safety as usage scales up.

Performance Metrics: A Comparative Analysis

When examining the technical capabilities of these advanced AI models, benchmark performance provides the most objective measure of their abilities. Gemini 2.5 Pro and GPT-4.5 each demonstrate unique strengths across various domains, with benchmark tests revealing their distinct advantages.

Gemini 2.5 Pro shows exceptional strength in reasoning-intensive tasks, particularly excelling in long-context reasoning and knowledge retention. It significantly outperforms competitors on Humanity’s Last Exam, which tests the frontier of human knowledge. However, it shows relative weaknesses in code generation, agentic coding, and occasionally struggles with factuality in certain domains.

GPT-4.5, conversely, demonstrates particular excellence in pattern recognition, creative insight generation, and scientific reasoning. It outperforms in the GPQA diamond benchmark, showing strong capabilities in scientific domains. The model also exhibits enhanced emotional intelligence and aesthetic intuition, making it particularly valuable for creative and design-oriented applications. A key advantage is its reduced tendency to generate false information compared to its predecessors.

In practical terms, Gemini 2.5 Pro represents the superior choice for tasks requiring deep reasoning, multimodal understanding, and handling extremely long contexts. GPT-4.5 offers advantages in creative work, design assistance, and applications where factual precision and natural conversational flow are paramount.

Applications and Use Cases

While benchmark performances provide valuable technical insights, the true measure of these advanced AI models lies in their practical applications across various domains. Both Gemini 2.5 Pro and GPT-4.5 demonstrate distinct strengths that make them suitable for different use cases, with organizations already beginning to leverage their capabilities to solve complex problems.

Gemini 2.5 Pro in Scientific and Technical Domains

Gemini 2.5 Pro’s exceptional reasoning capabilities and extensive context window make it particularly valuable for scientific research and technical applications. Its ability to process and analyze multimodal data—including text, images, audio, video, and code—enables it to handle complex problems that require synthesizing information from diverse sources. This versatility opens up numerous possibilities across industries requiring technical precision and comprehensive analysis.

1. Scientific research and data analysis – Gemini 2.5 Pro’s strong performance on benchmarks like GPQA (79.7%) demonstrates its potential to assist researchers in analyzing complex scientific literature, generating hypotheses, and interpreting experimental results
2. Software development and engineering – The model excels at creating web applications, performing code transformations, and developing complex programs with a 63.8% score on SWE-Bench Verified using custom agent setups
3. Medical diagnosis and healthcare – Its reasoning capabilities enable analysis of medical imagery alongside patient data to support healthcare professionals in diagnostic processes
4. Big data analytics and knowledge management – The 1 million token context window (expanding soon to 2 million) allows processing of entire datasets and code repositories in a single prompt

GPT-4.5’s Excellence in Creative and Communication Tasks

In contrast, GPT-4.5 demonstrates particular strength in tasks requiring nuanced communication, creative thinking, and aesthetic judgment. OpenAI emphasized training this model specifically for human collaboration, resulting in enhanced capabilities for content creation, design assistance, and natural communication.

1. Content creation and writing – GPT-4.5 shows enhanced aesthetic intuition and creativity, making it valuable for generating marketing copy, articles, scripts, and other written content
2. Design collaboration – The model’s improved understanding of nuance and context makes it an effective partner in design processes, from conceptualization to refinement
3. Customer engagement – With greater emotional intelligence, GPT-4.5 provides more appropriate and natural responses in customer service contexts
4. Educational content development – The model excels at tailoring explanations to different knowledge levels and learning styles

Companies across various sectors are already integrating these models into their workflows. Microsoft has incorporated OpenAI’s technology directly into its product suite, providing enterprise users with immediate access to GPT-4.5’s capabilities. Similarly, Google’s Gemini 2.5 Pro is finding applications in research institutions and technology companies seeking to leverage its reasoning and multimodal strengths.

The complementary strengths of these models suggest that many organizations may benefit from utilizing both, depending on specific use cases. As these technologies continue to mature, we can expect to see increasingly sophisticated applications that fundamentally transform knowledge work, creative processes, and problem-solving across industries.

The Future of AI: What’s Next?

As Gemini 2.5 Pro and GPT-4.5 push the boundaries of what’s possible, the future trajectory of AI development comes into sharper focus. Google’s commitment to “building thinking capabilities directly into all models” suggests a future where reasoning becomes standard across AI systems. Similarly, OpenAI’s approach of “scaling unsupervised learning and reasoning” points to models with ever-expanding capabilities to understand and generate human-like content.

The coming years will likely see AI models with dramatically expanded context windows beyond the current limits, more sophisticated reasoning, and seamless integration across all modalities. We may also witness the rise of truly autonomous AI agents capable of executing complex tasks with minimal human supervision. However, these advancements bring significant challenges. As AI capabilities increase, so too does the importance of addressing potential risks related to misinformation, privacy, and the displacement of human labor.

Ethical considerations must remain at the forefront of AI development. OpenAI acknowledges that “each increase in model capabilities is an opportunity to make models safer”, highlighting the dual responsibility of advancement and protection. The AI community will need to develop robust governance frameworks that encourage innovation while safeguarding against misuse.

The AI revolution represented by Gemini 2.5 Pro and GPT-4.5 is only beginning. While the pace of advancement brings both excitement and apprehension, one thing remains clear: the future of AI will be defined not just by technological capabilities, but by how we choose to harness them for human benefit. By prioritizing responsible development that augments human potential rather than replacing it, we can ensure that the next generation of AI models serve as powerful tools for collective progress.

The post Gemini 2.5 Pro and GPT-4.5: Who Leads the AI Revolution? appeared first on Cody - The AI Trained on Your Business.

OpenAI’s GPT-4.5, while a minor update, boasts improvements in reducing hallucinations and enhancing natural conversation. On the other hand, Anthropic’s Claude 3.7 Sonnet has garnered attention for its exceptional coding abilities and cost-effectiveness. Both models cater to a wide range of users, from developers and researchers to businesses seeking cutting-edge AI solutions.

As these models push the boundaries of what’s possible in AI, they’re reshaping expectations and applications across various industries, setting the stage for even more transformative advancements in the near future.

Key Features of GPT-4.5 and Claude 3.7 Sonnet

Both GPT-4.5 and Claude 3.7 Sonnet bring significant advancements to the AI landscape, each with its unique strengths. GPT-4.5, described as OpenAI’s “largest and most knowledgeable model yet,” focuses on expanding unsupervised learning to enhance word knowledge and intuition while reducing hallucinations. This model excels in improving reasoning capabilities and enhancing chat interactions with deeper contextual understanding.

On the other hand, Claude 3.7 Sonnet introduces a groundbreaking hybrid reasoning model, allowing for both quick responses and extended, step-by-step thinking. It particularly shines in coding and front-end web development, showcasing excellent instruction-following and general reasoning abilities.

Key Improvements:

• GPT-4.5: Enhanced unsupervised learning and conversational capabilities
• Claude 3.7 Sonnet: Advanced hybrid reasoning and superior coding prowess
• Both models: Improved multimodal capabilities and adaptive reasoning

Performance and Evaluation

Task	GPT-4.5 (vs 4o)	Claude 3.7 Sonnet* (vs 3.5)
Coding	Improved	Significantly outperforms
Math	Moderate improvement	Better on AIME’24 problems
Reasoning	Similar performance	Similar performance
Multimodal	Similar performance	Similar performance

* Without extended thinking

GPT-4.5 has shown notable improvements in chat interactions and reduced hallucinations. Human testers have evaluated it to be more accurate and factual compared to previous models, making it a more reliable conversational partner.

Claude 3.7 Sonnet, on the other hand, demonstrates exceptional efficiency in real-time applications and coding tasks. It has achieved state-of-the-art performance on SWE-bench Verified and TAU-bench, showcasing its prowess in software engineering and complex problem-solving. Additionally, its higher throughput compared to GPT-4.5 makes it particularly suitable for tasks requiring quick responses and processing large amounts of data.

Source: Anthropic

Pricing and Accessibility

GPT-4.5, while boasting impressive capabilities, comes with a hefty price tag. It’s priced 75 times higher than its predecessor, GPT-4, without clear justification for the substantial increase. This pricing strategy may limit its accessibility to many potential users.

In contrast, Claude 3.7 Sonnet offers a more affordable option. Its pricing structure is significantly more competitive:

1. 25 times cheaper for input tokens compared to GPT-4.5
2. 10 times cheaper for output tokens
3. Specific pricing: $3 per million input tokens and $15 per million output tokens

Regarding availability, GPT-4.5 is currently accessible to GPT Pro users and developers via API, with plans to extend access to Plus users, educational institutions, and teams. Claude 3.7 Sonnet, however, offers broader accessibility across all Claude plans (Free, Pro, Team, Enterprise), as well as through the Anthropic API, Amazon Bedrock, and Google Cloud’s Vertex AI.

These differences in pricing and accessibility significantly impact the potential adoption and use cases for each model, with Claude 3.7 Sonnet potentially appealing to a wider range of users due to its cost-effectiveness and broader availability.

Use Cases

Both GPT-4.5 and Claude 3.7 Sonnet offer unique capabilities that cater to diverse real-world applications. GPT-4.5 excels as an advanced conversational partner, surpassing previous models in accuracy and reducing hallucinations. Its improved contextual understanding makes it ideal for customer service, content creation, and personalized learning experiences.

Claude 3.7 Sonnet, on the other hand, shines in the realm of coding and software development. Its agentic coding capabilities, demonstrated through Claude Code, automate tasks like searching code, running tests, and using command line tools. This makes it an invaluable asset for businesses looking to streamline their development processes.

Future Prospects and Conclusion

The release of GPT-4.5 and Claude 3.7 Sonnet marks a significant milestone in AI development, setting the stage for even more groundbreaking advancements. While GPT-4.5 is seen as a minor update, it lays the foundation for future models with enhanced reasoning capabilities. Claude 3.7 Sonnet, with its hybrid reasoning model, represents a dynamic shift in the AI landscape, potentially influencing the direction of future developments.

As these models continue to evolve, we can anticipate further improvements in unsupervised learning, reasoning capabilities, and task-specific optimizations. The complementary nature of unsupervised learning and reasoning suggests that future AI models will likely exhibit even more sophisticated problem-solving abilities.

The post GPT-4.5 vs Claude 3.7 Sonnet: A Deep Dive into AI Advancements appeared first on Cody - The AI Trained on Your Business.

The post Perplexity Comet: Bold Leap into Agentic Search appeared first on Cody - The AI Trained on Your Business.