OpenAI Unveils Budget-Friendly GPT-4o Mini, Which Slashes Costs By 97%
AInvestThursday, Jul 18, 2024 11:23 pm ET
2min read
EFSC --
GOOGL --
META --

On Thursday, OpenAI announced the official launch of a new, significantly cheaper entry-level AI model, GPT-4o mini.

In May this year, OpenAI released GPT-4o, the fastest, most comprehensive, and also the most expensive model in the company's history. The newly launched GPT-4o mini is a smaller, more affordable variant, designed to compete in the lower-priced market by balancing capability and cost-effectiveness.

According to OpenAI, the API price of GPT-4o mini will be 15 cents per million input tokens and 60 cents per million output tokens. The context window remains at 128,000 tokens, with a knowledge cutoff in October 2023. The company declined to disclose the parameter size of the new model, stating only that it is in the same league as similar small models on the market (such as Claude Haiku and Gemini 1.5 Flash).

By comparison, the price for GPT-4o is $5/$15 per million input/output tokens, while the previous entry-level model, GPT-3.5 Turbo, is priced at 50 cents/$1.5. Thus, the latest mini model is 96%-97% cheaper than GPT-4o and 60%-70% cheaper than GPT-3.5 Turbo. With the launch of the GPT-4o mini, the GPT-3.5 Turbo has reached the end of its historical mission.

According to statistics from Artificial Analysis, the GPT-4o mini has reached the lowest price point among mainstream small models from major American AI companies, being significantly cheaper than similar products from Anthropic and Google. However, even OpenAI cannot compete with the pricing of Eastern forces. For instance, the token output price of the DeepSeek-V2 model by the well-known private equity firm Phantom is still half the price of the GPT-4o mini.

In terms of capabilities, the GPT-4o mini is also quite competitive. OpenAI mentioned that in the MMLU (Massive Multitask Language Understanding) test, GPT-4o mini scored 82%. Currently, the highest score in this test is held by Google's ultra-large model Gemini Ultra-1760B (90%), with GPT-4o scoring 88.7%.

Among the small models, Gemini 1.5 Flash scored 79%, and Claude 3 Haiku scored 75%. Although the recognition of large model test benchmarks varies, this comparison roughly indicates the positioning of GPT-4o mini.

OpenAI stated that GPT-4o mini can initially handle text and image generation tasks, to eventually support all other types of content processing. Starting Thursday, OpenAI will gradually roll out this new model to free and paid ChatGPT users, replacing GPT-3.5 Turbo. Enterprise users will be able to access the new product next week.

There is also a noteworthy security update associated with this new model. OpenAI revealed that GPT-4o mini is the company's first AI model to employ a new security strategy known as instruction hierarchy. This strategy requires the AI system to prioritize certain instructions - such as preset commands from OpenAI - making it more difficult for malicious users to have the tool perform undesirable operations.

Although OpenAI claims to make AI accessible to as many people as possible, the pricing reflects the fierce competition among AI cloud service providers.

OpenAI Product Manager Olivier Godement explained that the lower pricing is due to improvements in model architecture, optimized training data, and training schedules. However, he also emphasized that the GPT-4o mini outperforms similar small models in common market benchmarks.

For OpenAI, the largest challenge will come next week. According to multiple sources, Meta is set to release the largest parameter version of its open-source large model, Llama 3, on July 23, reportedly boasting 400 billion parameters. It remains unclear how powerful this model will be.

Disclaimer: the above is a summary showing certain market information. AInvest is not responsible for any data errors, omissions or other information that may be displayed incorrectly as the data is derived from a third party source. Communications displaying market prices, data and other information available in this post are meant for informational purposes only and are not intended as an offer or solicitation for the purchase or sale of any security. Please do your own research when investing. All investments involve risk and the past performance of a security, or financial product does not guarantee future results or returns. Keep in mind that while diversification may help spread risk, it does not assure a profit, or protect against loss in a down market.