Exploring the New Frontier of AI: OpenAI’s GPT-4-o For Indic Languages

In the ever-evolving landscape of artificial intelligence, OpenAI has once again pushed the boundaries with the introduction of the GPT-4-o model, featuring the innovative o200k_base tokenizer. This development marks a significant leap forward in the field, offering unprecedented speed, affordability, and multimodal capabilities.

What is GPT-4-o?

GPT-4-o, where the ‘o' stands for “omni,” is OpenAI's latest flagship generative model introduced on May 13, 2024. It is designed to handle a diverse array of inputs including text, speech, and video, and can generate outputs in various formats such as text, audio, and images. This versatility makes it a powerful tool for a wide range of applications.  This integration marks a pivotal evolution from its predecessors, primarily focusing on text-based processing.

The o200k_base Tokenizer

The o200k_base tokenizer is a new tokenization algorithm that forms the backbone of the GPT-4-o model. Tokenization is a critical process in natural language processing that involves breaking down text into smaller units called tokens. These tokens can be words, subwords, or even characters, depending on the tokenizer's design.

The o200k_base tokenizer represents an evolution in this process, designed to be faster and more efficient than its predecessors. It allows GPT-4-o to process and generate language at speeds that were previously unattainable. The o200kbase tokenizer not only improves the semantic coherence of the generated text but also plays a crucial role in handling multiple languages more effectively, thereby broadening the scope of GPT-4o's applications across different linguistic contexts.

Features and Capabilities

  • Multimodal Inputs and Outputs: GPT-4-o accepts and emits a variety of data types, setting it apart from earlier models that were limited to text. This makes it an “omni” model, capable of more complex tasks that mirror human interaction with various forms of data.
  • Improved Token Generation Speed: GPT-4-o is reported to generate tokens twice as fast as GPT-4 Turbo, enhancing its efficiency and making it suitable for real-time applications.
  • Cost-Effectiveness: Despite its advanced capabilities, GPT-4-o is more affordable than its predecessors. The API costs have been significantly reduced, making it accessible for a broader range of users and developers.
  • Enhanced Vision Capabilities: Compared to previous models, GPT-4-o has improved vision capabilities, allowing it to handle tasks involving image recognition and manipulation with greater finesse.

Analysis of Indic languages

The analysis of the o200k_base tokenizer's performance across various Indic languages demonstrates significant improvements in efficiency and reduction in token usage when working with GPT-4o models. The data highlights that the Malayalam language experienced the most substantial efficiency improvement of almost 4x. Kannada and Telugu also show impressive improvements, with reduction percentages nearing 79% and 77%, respectively, and high improvement factors suggesting much greater processing efficiency. This trend continues notably with Gujarati and Tamil, showcasing over 74% reduction in token usage. On the lower end of the scale, languages like Kashmiri and Manipuri displayed lesser improvement, with Kashmiri only showing a 37.70% reduction and Manipuri showing no improvement in token usage. This indicates variability in how the new tokenizer handles different linguistic structures and , which might be due to the inherent linguistic features or the training data's coverage and quality. 


Language Name Avg Tokens GPT-4 Avg Tokens GPT-4o Avg % Reduction in Tokens Improvement Factor
Malayalam (മലയാളം) 4775 957 79.35% 3.99x
Kannada (ಕನ್ನಡ) 3681 766 78.83% 3.8x
Telugu (తెలుగు) 4097 893 76.63% 3.59x
Gujarati (ગુજરાતી) 3408 758 74.36% 3.49x
Tamil (தமிழ்) 3949 948 74.46% 3.17x
Bangla (বাংলা) 2550 704 70.06% 2.62x
Punjabi (ਪੰਜਾਬੀ) 4208 1297 67.73% 2.24x
Assamese (অসমীয়া) 2866 884 67.11% 2.24x
Hindi (हिन्दी) 2090 655 64.20% 2.19x
Nepali (नेपाली) 2638 878 61.59% 2.0x
Urdu (اردو) 2428 854 62.31% 1.84x
Marathi (मराठी) 2593 912 62.65% 1.84x
Bhojpuri (Bhojpuri) 1970 699 62.31% 1.82x
Chhattisgarhi  1958 733 59.89% 1.67x
Maithili (Maithili) 1975 767 60.04% 1.58x
Odia (ଓଡ଼ିଆ) 6074 2432 60.34% 1.5x
Konkani (Konkani) 2135 875 56.91% 1.44x
Sindhi (سنڌي) 2188 921 55.08% 1.37x
Dogri (Dogri) 2361 1025 55.98% 1.3x
Kashmiri (کٲشُر) 2291 1484 37.70% 0.54x
Manipuri  6715 6715 0.00% 0.0x



Now we if we look from a cost perspective we get the additional benefit as GPT-4o is offered in 50% reduction in pricing compared to GPT-4-Turbo which then leads to further reduction in overall cost of typical RAG request. Here is a comparison of a typical RAG request with 1000 input words and 200 output words. Over all there is almost 5 fold reduction in overall cost.


How did we analyze?

The analysis of the o200k_base tokenizer's performance across Indic languages was meticulously conducted using English language documents of varying lengths—approximately 10, 100, 500, and 1200 words. These documents were translated into each target Indic language using Azure Translator. Each translated document was then processed through both the tokenizer for GPT-4 and GPT-4o models to assess and record the number of tokens required by each model. This method allowed us to compare the efficiency of the new o200k_base tokenizer against its predecessor across different text lengths, providing a broad and robust dataset for analysis. After processing, the token counts from each document size were averaged to mitigate any anomalies that might occur at specific text lengths and to provide a more generalized view of performance across typical usage scenarios. 

Real-world Applications

The implications of GPT-4-o's capabilities are vast. Here are just a few potential applications:

  • Language Translation: With its efficient tokenization, GPT-4-o could provide near-instantaneous translation across multiple languages, breaking down communication barriers.
  • Content Creation: The model's ability to handle text and images makes it an excellent tool for content creators, enabling the generation of rich multimedia content.
  • Educational Tools: GPT-4-o could revolutionize online learning by providing interactive multimodal content that adapts to various learning styles.
  • Accessibility Features: The model can convert speech to text and vice versa, offering new tools for individuals with disabilities to interact with technology.


The GPT-4-o model with the o200k_base tokenizer is a testament to OpenAI's commitment to advancing technology. By enhancing speed, reducing costs, and expanding capabilities, GPT-4-o stands to democratize access to cutting-edge tools and pave the way for innovative applications that were once the realm of science fiction. As we stand on the brink of this new AI era, it is clear that OpenAI's GPT-4-o is not just a technological milestone but also a harbinger of a future where AI and human creativity converge in exciting and transformative ways.


This article was originally published by Microsoft's Azure AI Services Blog. You can find the original article here.