GPT-4.5 for enterprise: Do its accuracy and knowledge justify the cost?

Release Openai GPT-4.5 slightly disappointed, many showing the crazy price point (about 10 to 20x) Claude 3.7 Sonnet and more expensive than 15 and 30x GPT-4O).

However, given that this is the biggest and most powerful model of this, it is worth considering its strong and glowing areas.

Better knowledge and adaptation

There is little detail about the architecture or exercise of the model, but we have a rough assessment taught by 10x more calculations. And the model was so large that a lot of information are spread to distribute training in many data center over a period of time.

In large models, there is a greater potential to learn world knowledge and the nuances of the human language (taking into account high quality training information). This is known in some sizes presented by Openai team. For example, GPT-4.5 has a record high rank on a benchmark evaluating hallucinations in AI models.

Practical practices also show that GPT-4.5 is better than other common model models remaining in accordance with facts and user instructions.

Users noted that GPT-4.5 answers are more natural and aware of the context than previous models. The ability to watch tons and style rules also improved.

GPT-4.5, AI scientist and Openai co-founder Andrej Karpaty, the model came out early, gossip he is “waiting[ed] Improving the basic tasks, I would say that these are more related to EQ (unlike IQ) and for example, world knowledge, similar, analogy, general understanding, humor, etc.

However, the quality of writing is also very subjective. In a study, Carpathy fled to different instructions, most people preferred the answers to GPT-4O over GPT-4.5. O wrote in x: “Highly delicious tests focus on a new and unique structure, but low tastes are overcome. Or we just work. Or these examples are not just so big. Or actually is quite close and this is a very small example size. Or all of the above. “

Better document processing

In their experiments, in the box containing the box Integrated GPT-4.5 This is one of the best models in the Mission of GPT-4.5 “GPT-4.5”, accuracy and integrity to the AI ​​studio product, but also the most difficult AI questions we encounter are one of the best models to manage.

In domestic evaluations, the box has found GPT-4.5 more accurate about the question-answer tasks of the Enterprise document – learn the most with the original GPT-4 point in the test set.

Box tests also showed that GPT-4.5 has prevailed in the mathematical questions of GPT-4.5 in their mathematical questions, which are often struggling with older GPT models. For example, it was better to answer questions about financial documents that require the implementation of information and calculations.

GPT-4.5 also showed improved performance in extracting information from unstructured data. In a test that produces areas of hundreds of legal documents, GPT-4.5 GPT-4O was 19% more accurate.

Planning, coding, results evaluation

Given the improved world knowledge, GPT-4.5 can also be a suitable model to create high-level plans for complex tasks. After that, the broken steps will be handed over and executed in smaller but more efficient models.

According to Zodiac survey“An agentic planning and execution of agent, including the initial test, GPT-4.5, multi-step coding work flows and complex task automation, show strong opportunities.”

GPT-4.5 can also be useful in coding tasks that require internal and contextual knowledge. GitHub now provides Limited Input The Copilot implements the model and GPT-4.5 “creative proposals in the coding assistant and provides valid responses to dark knowledge inquiries.”

Given the knowledge of the deeper world, GPT-4.5 is also suitable for “LLM-AS-A-HAK“Tasks assessed by a strong model of small models. For example, a model like GPT-4O or O3 can result in one or more answers, and transmit the final answer to GPT-4.5 for revision and elegance.

Is it worth the price?

Given the great costs of GPT-4.5, it is very difficult to justify many of the events used. But this does not mean that it will remain in this way. One Sustainable trends In recent years, we saw that if this trend belongs to GPT-4.5, the trend is worth finding the way to practice with GPT-4.5 and to use it in enterprise applications.

It should be noted that this new model can be the basis for future mediocre models. Per Carpathy: “Keep in mind that GPT4.5 is only preterraining, controlled finetuning and RLHF [reinforcement learning from human feedback]Therefore, this is not yet a substantial model. Therefore, in cases where this model of release is critical (mathematics, code, etc.), the model is likely to think that Openai thinks and push the model in these domains. “