Claude 3.5 Sonnet vs. GPT-4o: here’s how they stack up

ChatGPT users can access the GPT-4o model for free, but the number of queries you’re allowed to ask it within a three-hour span is limited. Exceeding that limit will move you down to using GPT-3.5, OpenAI’s older, less capable model, until the cooldown timer has reset.

GPT-4o, on the other hand, offers users a wider variety of functions, a larger knowledge base, and more flexible log-in options than Claude.

Unlike Anthropic’s offering, ChatGPT can generate images (thoughthat feature requires a subscription) and access the web to find the most up-to-date information available, not just the data it was trained on. And contrary to the sign-up process with Claude, GPT-4o users won’t have to provide their phone number for verification, or really any credentials at all, in order to start using the model And, thanks to its larger knowledge base, GPT-4o is more adept at generating creative content across images, text, and audio than Claude.

The ChatGPT website also tends to overload with user requests at a far lower rate than Claude’s.

Both models are capable of generating a wide variety of text outputs, from refining user-written prose to creating computer code and spreadsheet algorithms. When asked to generate a “piece of Python code that will measure the area of a circle and generate a graph of the function,” both returned similarly structured snippets. GPT-4o produced the following:

import numpy as np
import matplotlib.pyplot as plt

# Function to calculate the area of a circle
def calculate_area(radius):
return np.pi * radius ** 2

# Generate a range of radius values
radii = np.linspace(0, 10, 100) # radii from 0 to 10 with 100 points

# Calculate the corresponding area values
areas = calculate_area(radii)

# Plotting the function
plt.figure(figsize=(10, 6))
plt.plot(radii, areas, label=’Area = πr²’)
plt.title(‘Area of a Circle as a Function of Radius’)
plt.xlabel(‘Radius’)
plt.ylabel(‘Area’)
plt.legend()
plt.grid(True)
plt.show()

Meanwhile, Claude’s response was a bit more fleshed out:

import math
import numpy as np
import matplotlib.pyplot as plt

def circle_area(radius):
return math.pi * radius**2

# Generate radii from 0 to 10 with 100 points
radii = np.linspace(0, 10, 100)

# Calculate areas for each radius
areas = [circle_area(r) for r in radii]

# Create the plot
plt.figure(figsize=(10, 6))
plt.plot(radii, areas, ‘b-‘, linewidth=2)
plt.title(‘Area of a Circle as a Function of Radius’)
plt.xlabel(‘Radius’)
plt.ylabel(‘Area’)
plt.grid(True)

# Add some interesting points
plt.plot(0, circle_area(0), ‘ro’, label=’r = 0′)
plt.plot(1, circle_area(1), ‘ro’, label=’r = 1′)
plt.plot(5, circle_area(5), ‘ro’, label=’r = 5′)
plt.plot(10, circle_area(10), ‘ro’, label=’r = 10′)

plt.legend()
plt.show()

In terms of image identification, the two models are closely matched. When asked to identify the various aspects of the following image, both GPT-4o and Claude 3.5 were easily able to find and describe each of the 21 items present.

GPT-4o has the advantage of delivering information about recent events, given its ability to search the web for news that’s more recent than its training data. Asking Claude “what happened in Luxembourg yesterday” only gets you a note reading: “I apologize, but I don’t have access to real-time news or information about specific events that occurred yesterday in Luxembourg. My knowledge cutoff is in April 2024, and I don’t have information about events after that date.” The AI does offer a number of helpful recommendations as to where you might be able to find the information you requested.

I did find comparing the two systems to be a bit of a challenge, mostly because I was only able to make a half-dozen requests of Claude before the system locked me out for three hours. Even on the free tier of ChatGPT, I won’t be frozen out of the system entirely — I’ll just have to converse with a slightly inferior model for a while.