Concerns about TPM Limitations in GPT-4 Turbo for Chatbot Applications

Hello, folks. I am seeing TPM (Token per minutes) from ‘Rate Limits’.
image

Our company is using a chatbot, and not only that, we are planning to use this for some other tasks as well, along with the whisper-1 call.

Now, what I am worried about is that we have some TPM. For example, gpt-4-turbo-2024-04-09 is showing ONLY 25,000 TPM, but turbo-preview was granting 1,500,000 TPM. This is actually a huge difference, such that our team cannot use gpt-4-turbo-2024-04-09 for our use cases.

I do not believe that 25,000 TPM is sufficient to handle chats, considering that OpenAI was permitting over 8k tokens to be fed.

Are there any plans to increase the capacity?

250.000 TPM btw.

do you need vision for all your calls? use turbo-preview where you don’t need vision, and turbo where you do :thinking:

for me it’s showing this: (t3)

it’s possible that gpt-4-turbo and and gpt-4-turbo-2024-04-09 might not share a limit, even though they point to the same model :thinking:

1 Like

That is obviously a brand new model just introduced, now with no more “preview”.

gpt-4-turbo-2024-04-09 GPT-4 Turbo with Vision model. – Vision requests can now use JSON mode and function calling. gpt-4-turbo currently points to this version.

Vision was previously based on 1106.

The 250k rate for it, across all tiers, is the “all other models” rate. It will probably be set to a real rate shortly.

1 Like

Thank you for your quick response.

I see… it was a vision model. I should have looked at the models link first. My bad. Models - OpenAI API

I was confused before I looked at this API. “gpt-4-vision-preview” is for vision, so I thought the model name would be something like “gpt-4-vision-2024-04-09” for vision. But yeah… again, my bad.

I hope they are not planning to remove the higher TPM models with GPT-4.

Everything is fine now. :slight_smile:

2 Likes

dude, don’t even worry about it. OpenAI naming is… …not exactly following any industry standard I know of, I guess. :laughing:

1 Like

The model being named “gpt-4-turbo” likely means it has vision, but is to be a general-purpose model, doing what ChatGPT has been doing for half a year.

Just had an investigation: logprobs work, logit_bias works. Very fast compared to a stuttering “vision”, likely because nobody is using this yet.


Compared to 1106 based vision, this new “fast” gives a vision answer even more wrong – It doesn’t even understand the desired answer.

image

gpt-4-vision-preview:

To solve the depicted geometry problem, we will follow these steps:

  1. Understand the problem: We have two circles, A and B, where the radius of circle A is one-third the radius of circle B. Circle A rolls around circle B without slipping.

  2. Determine the circumference of each circle: The circumference of a circle is given by the formula ( C = 2\pi r ), where ( r ) is the radius of the circle. Since the radius of circle A is one-third of circle B, the circumference of circle A will also be one-third of the circumference of circle B.

  3. Calculate the distance circle A travels in one revolution around circle B: When circle A completes one revolution around circle B, it will have traveled a distance equal to the circumference of circle B.

  4. Determine the number of revolutions of circle A needed to travel the circumference of circle B: Since the circumference of circle A is one-third of circle B, circle A will need to complete three revolutions to travel a distance equal to the circumference of circle B.

  5. Find the point of return to the starting position: We need to find out after how many revolutions of circle A its center will return to its starting point on circle B.

Now, let’s solve the problem using these steps:

  1. Let the radius of circle B be ( r_B ) and the radius of circle A be ( r_A ). According to the problem, ( r_A = \frac{1}{3}r_B ).

  2. The circumference of circle B, ( C_B ), is ( 2\pi r_B ), and the circumference of circle A, ( C_A ), is ( 2\pi r_A = 2\pi \left(\frac{1}{3}r_B\right) = \frac{2\pi r_B}{3} ).

  3. When circle A rolls around circle B once, it travels a distance equal to ( C_B ), which is ( 2\pi r_B ).

  4. To travel ( 2\pi r_B ), circle A must revolve ( \frac{2\pi r_B}{C_A} ) times. Since ( C_A = \frac{2\pi r_B}{3} ), the number of revolutions of circle A is ( \frac{2\pi r_B}{\frac{2\pi r_B}{3}} = 3 ).

  5. Therefore, circle A will return to its starting point after 3 revolutions around circle B.

gpt-4-turbo

Step-by-Step Overview of the Problem

  1. Understanding the Geometry: The problem involves two circles, A and B, where circle A has a radius that is 1/3 the radius of circle B. Circle A rolls around circle B without slipping.

  2. Concept of Rolling Without Slipping: When a circle rolls without slipping, the distance it covers along the circumference of the other circle equals the arc length it rolls through on its own circumference.

  3. Determine the Circumferences:

    • The circumference of circle A (C_A) is (2\pi r_A).
    • The circumference of circle B (C_B) is (2\pi r_B).
  4. Relate the Radii and Circumferences: Given (r_A = \frac{1}{3}r_B), it follows that (C_A = \frac{1}{3}C_B).

  5. Calculate Revolutions of Circle A Around Circle B: To find out how many revolutions circle A makes around circle B before its center returns to the starting point, we need to consider the total distance circle A travels as it rolls around circle B and how this relates to its own circumference.

Approach to Solving the Problem

  • Calculate the Total Distance Circle A Travels: As circle A rolls around circle B, the distance it travels is the circumference of circle B, (C_B).
  • Number of Revolutions of Circle A on Its Own Axis: Since (C_A = \frac{1}{3}C_B), circle A will make three complete revolutions on its own axis for every one complete revolution around circle B.
  • Return to Starting Point: We need to determine how many times circle A needs to revolve around circle B such that the center of circle A returns to its original position relative to circle B. This involves considering the relative motion of the centers of the two circles.

Solving the Problem

  1. Circumference of Circle B: (C_B = 2\pi r_B).
  2. Circumference of Circle A: (C_A = 2\pi r_A = 2\pi \left(\frac{1}{3}r_B\right) = \frac{2\pi r_B}{3} = \frac{1}{3}C_B).
  3. Revolutions of Circle A on Its Own Axis: For each complete revolution around circle B, circle A rolls through a distance of (C_B), which equals three revolutions of circle A on its own axis.
  4. Return to Starting Point: Since the radius of circle A is 1/3 of that of circle B, and it makes three revolutions on its own axis per one revolution around circle B, the center of circle A will align with its starting point after one complete revolution around circle B.

Conclusion

Circle A will return its center to the starting point after one complete revolution around circle B.

Both with ChatGPT TeX markdown OpenAI hasn’t given us the courtesy of documenting.

1 Like

By the way, this is just for the update:

Now I can see,

that gpt-4-turbo-2024-04-09 is now using the shared limits. :slight_smile:

So, it looks like, if I am fully ready, I can upgrade my current model, gpt-4-0125-preview, to gpt-4-turbo-2024-04-09, as they are showing the same price per token.

Pricing (openai.com)

Also, it appears that they have combined the -vision-preview and -preview models.

Perfect.

Thanks, it looks like you were correct, buddy.

gpt-4-turbo is now tiered - and giving tier-1 five times more TPM spending (and 50x more profit per org per minute) than gpt-3.5-turbo.

I guess that depends on your account provisioning… :face_with_spiral_eyes:

image