Beyond benchmarks: How DeepSeek-R1 and o1 perform on real-world tasks

DeepSeek-R1 has surely created a lot of excitement and concern, especially for OpenAI’s rival model o1. So, we put them to test in a side-by-side comparison on a few simple data analysis and market research tasks. 

To put the models on equal footing, we used Perplexity Pro Search, which now supports both o1 and R1. Our goal was to look beyond benchmarks and see if the models can actually perform ad hoc tasks that require gathering information from the web, picking out the right pieces of data and performing simple tasks that would require substantial manual effort. 

Both models are impressive but make errors when the prompts lack specificity. o1 is slightly better at reasoning tasks but R1’s transparency gives it an edge in cases (and there will be quite a few) where it makes mistakes.

Here is a breakdown of a few of our experiments and the links to the Perplexity pages where you can review the results yourself.

Calculating returns on investments from the web

Our first test gauged whether models could calculate returns on investment (ROI). We considered a scenario where the user has invested $140 in the Magnificent Seven (Alphabet, Amazon, Apple, Meta, Microsoft, Nvidia, Tesla) on the first day of every month from January to December 2024. We asked the model to calculate the value of the portfolio at the current date.

To accomplish this task, the model would have to pull Mag 7 price information for the first day of each month, split the monthly investment evenly across the stocks ($20 per stock), sum them up and calculate the portfolio value according to the value of the stocks on the current date.

In this task, both models failed. o1 returned a list of stock prices for January 2024 and January 2025 along with a formula to calculate the portfolio value. However, it failed to calculate the correct values and basically said that there would be no ROI. On the other hand, R1 made the mistake of only investing in January 2024 and calculating the returns for January 2025.

However, what was interesting was the models’ reasoning process. While o1 did not provide much details on how it had reached its results, R1’s reasoning traced showed that it did not have the correct information because Perplexity’s retrieval engine had failed to obtain the monthly data for stock prices (many retrieval-augmented generation applications fail not because of the model lack of abilities but because of bad retrieval). This proved to be an important bit of feedback that led us to the next experiment.

Reasoning over file content

We decided to run the same experiment as before, but instead of prompting the model to retrieve the information from the web, we decided to provide it in a text file. For this, we copy-pasted stock monthly data for each stock from Yahoo! Finance into a text file and gave it to the model. The file contained the name of each stock plus the HTML table that contained the price for the first day of each month from January to December 2024 and the last recorded price. The data was not cleaned to reduce the manual effort and test whether the model could pick the right parts from the data.

Again, both models failed to provide the right answer. o1 seemed to have extracted the data from the file, but suggested the calculation be done manually in a tool like Excel. The reasoning trace was very vague and did not contain any useful information to troubleshoot the model. R1 also failed and didn’t provide an answer, but the reasoning trace contained a lot of useful information.

For example, it was clear that the model had correctly parsed the HTML data for each stock and was able to extract the correct information. It had also been able to do the month-by-month calculation of investments, sum them and calculate the final value according to the latest stock price in the table. However, that final value remained in its reasoning chain and failed to make it into the final answer. The model had also been confounded by a row in the Nvidia chart that had marked the company’s 10:1 stock split on June 10, 2024, and ended up miscalculating the final value of the portfolio. 

Again, the real differentiator was not the result itself, but the ability to investigate how the model arrived at its response. In this case, R1 provided us with a better experience, allowing us to understand the model’s limitations and how we can reformulate our prompt and format our data to get better results in the future.

Comparing data over the web

Another experiment we carried out required the model to compare the stats of four leading NBA centers and determine which one had the best improvement in field goal percentage (FG%) from the 2022/2023 to the 2023/2024 seasons. This task required the model to do multi-step reasoning over different data points. The catch in the prompt was that it included Victor Wembanyama, who just entered the league as a rookie in 2023.

The retrieval for this prompt was much easier, since player stats are widely reported on the web and are usually included in their Wikipedia and NBA profiles. Both models answered correctly (it’s Giannis in case you were curious), although depending on the sources they used, their figures were a bit different. However, they did not realize that Wemby did not qualify for the comparison and gathered other stats from his time in the European league.

In its answer, R1 provided a better breakdown of the results with a comparison table along with links to the sources it used for its answer. The added context enabled us to correct the prompt. After we modified the prompt specifying that we were looking for FG% from NBA seasons, the model correctly ruled out Wemby from the results.

Final verdict

Reasoning models are powerful tools, but still have a ways to go before they can be fully trusted with tasks, especially as other components of large language model (LLM) applications continue to evolve. From our experiments, both o1 and R1 can still make basic mistakes. Despite showing impressive results, they still need a bit of handholding to give accurate results.

Ideally, a reasoning model should be able to explain to the user when it lacks information for the task. Alternatively, the reasoning trace of the model should be able to guide users to better understand mistakes and correct their prompts to increase the accuracy and stability of the model’s responses. In this regard, R1 had the upper hand. Hopefully, future reasoning models, including OpenAI’s upcoming o3 series, will provide users with more visibility and control.