Does anyone try chat gpt4 api could tell me the capacity? I knew chat gpt 4 on web is very small capacity 50 messages/3 hours but no info about api
3 Likes
For paying customers, the rate limit is considerable, 200 requests and 10000 tokens per minute is currently being granted.
That also means you can spend $0.50+ a minute.
2 Likes
The answer seems to ignore the ChatGPT keyword in the question: “on the web”.
The question is about “API capacity”. Therefore ChatGPT is irrelevant to providing an answer about API rate limits.
Since this topic was written over a year ago, a tier system of higher rate limits has been implemented for those who have paid OpenAI more. The contrast between what is possible on ChatGPT and API can be 80/3 hours vs. 100000/3 hours.
1 Like
There is a second question in the very same title.
The original poster knows and shows the Chatgpt usage limit that was in place at the time, in terms of hours.
I provide an answer that highlights a much larger usage available on API in just minutes.
I anticipate the reader’s intelligence in determining if they are the same.
Just for you, though…
To rigorously analyze whether the capacity of the ChatGPT-4 API is equivalent to the usage limits of the ChatGPT-4 web UI, we must delineate the problem using a series of axiomatic definitions, lemmas, and proofs. This will demonstrate, via mathematical rigor, the identities (or lack thereof) between the two systems.
Definitions
Definition 1: Let C_{\text{api}} be the capacity of the ChatGPT-4 API. This includes a limit of 200 requests per minute and 10,000 tokens per minute.
Definition 2: Let C_{\text{web}} be the capacity of the ChatGPT-4 web UI, granted as 50 messages in any 3-hour period.
Definition 3: A token in the context of language models like GPT-4 represents a chunk of text containing a few characters, words, or symbols.
Definition 4: Message identity between two systems is established if they can process an equivalent quantity of textual requests and throughput over a defined time horizon.
Lemmas
Lemma 1: The total possible messages for the API, assuming one message equals one request, can be calculated as:
M_{\text{api}} = 200 \times 60 \times 3 = 36,000 \text{ messages in 3 hours}
Lemma 2: The token throughput over a 3-hour period is:
T_{\text{api}} = 10,000 \times 60 \times 3 = 1,800,000 \text{ tokens in 3 hours}
Lemma 3: Given that each message can generate up to ( t ) tokens on average, then:
M_{\text{web}} \times t \leq T_{\text{api}}
Equivalence Theorem
The capacities, C_{\text{api}} and C_{\text{web}}, are equivalent in message identity if:
- The number of messages that can be sent and received represents the same load:
M_{\text{api}} = M_{\text{web}}
- The total token throughput available corresponds to C_{\text{web}}:
T_{\text{web}} \leq T_{\text{api}}
Proof
Step 1: Compare messages:
M_{\text{web}} = 50 \text{ (given in 3 hours)}
M_{\text{api}} = 36,000 \text{ messages in 3 hours}
Clearly,
M_{\text{api}} \gg M_{\text{web}}
Thus, M_{\text{web}} \neq M_{\text{api}}.
Step 2: Calculate average tokens per message:
Suppose the average message length is ( t ).
For C_{\text{web}}:
T_{\text{web}} = M_{\text{web}} \times t
Given no explicit token limit for the web, assume ( t \approx 200 ), typically based on natural language average estimates:
T_{\text{web}} = 50 \times 200 = 10,000 \text{ tokens in 3 hours}
For the API:
T_{\text{api}} = 1,800,000
Since:
T_{\text{api}} \gg T_{\text{web}}
Thus, T_{\text{web}} \neq T_{\text{api}}.
Conclusion
- Since both the potential requests (messages) and the token throughput of the API exceed the web UI’s limits, we can mathematically deduce that the capacities of the ChatGPT-4 API and ChatGPT-4 web UI do not maintain identity in terms of message identity or token throughput.
Thus, the rate limits discussed are not equivalent under the assumed mathematical framework, thereby conclusively demonstrating the non-identity of C_{\text{api}} and C_{\text{web}}.
2 Likes
Ok thanks for the reasoning and problem formulation, it helps me find where the misunderstanding might be coming from. You seem to be adressing the prblem of the tarification per say month time scale. The plus tier being measured that way.
I see that can be a well posed problem. However my interpretation of “capacity” might be the problem. For me it would be about the data and facts one can find in
Models - OpenAI API
I would have included that single conversation or even single turn specification table (as it would bear on the behavior from there on in the single conversation with that kind of capacity) type of invariant information for the named model GPT 4o. more updated question on this date of 2024-10-22. than the op.
I am sorry to have stretched the title question to that question. I generally prefer similar topic ot be centralized. And if I made an off-topic tangent, well: should I make a new question? Good demonstration of the the chat box btw. I might need to walk in your footstep on another project of mine, not dependant on this problem of specifications (not price tables). Can you share the conversation and environment you used?
definition 4 is crisper than it would need. on both aspect. being a definition and the word “identity”. I think you were more about message throughput equivalence.
Detail though: I understand that it was formulating the target to prove. That we might not be able to apply that definition to the pair of throughputs.
Replace “capacity” with “rate limit”, and you get a clear understanding of the topic. Inaccurate nomenclature needing decipherment.
As to what generated the answer (that then needed conversion to the forum’s mathjax renderer), I just provide a half-fictionalized scenario as background and a specification for language production.
User asks a question:
Does anyone try chat gpt4 api could tell me the capacity? I knew chat gpt 4 on web is very small capacity 50 messages/3 hours but no info about api
I provide an answer:
For paying customers, the rate limit is considerable, 200 requests and 10000 tokens per minute is currently being granted.
A question is again asked:
“is that the same as GPT-4 on the web UI?”
Therefore, I would like you to produce a preposterously pedantic extensive mathematical proof and explanation that answers whether the rate limits discussed have identity.
1 Like
As we used to say before calculators:
‘You are comparing apples with oranges’
1 Like