← All teardowns

OpenAI Platform · Request logs

OpenAI put frontier models behind one clean API. This rebuild keeps the log's exact structure — and makes a 429 at 9:41 impossible to miss.

$ tail -f requests --window 24hLiveAll modelsErrors only
18,340 requests · 3 errors
TimeRequestModelTokens inTokens outLatencyStatus
09:42:18
req_9f2c81ab04d1
/v1/chat/completions
gpt-52,8149123.42s200 OK
09:42:11
req_9f2c7d55e2b8
/v1/chat/completions
gpt-5-mini640188780ms200 OK
09:41:57
req_9f2c76c9aa03
/v1/embeddings
text-embedding-3-large5,120240ms200 OK
09:41:44
req_9f2c71124f77
/v1/chat/completions
gpt-53,90260ms429 Rate limit
09:41:31
req_9f2c6b8d31c5
/v1/chat/completions
o3-mini1,2042,0485.91s200 OK
09:41:02
req_9f2c64f0b9e2
/v1/chat/completions
gpt-5-mini51296610ms200 OK
09:40:48
req_9f2c5fa7d044
/v1/responses
gpt-58,1101,7447.28s200 OK
09:40:20
req_9f2c58c2e6a9
/v1/chat/completions
gpt-4.1990130ms400 Error
09:39:55
req_9f2c52091bd7
/v1/chat/completions
gpt-52,20145ms500 Error
Showing 9 of 18,3401 rate-limited2 failed

Rebuilt by Kyre — same layout, same data, same information architecture. Only the execution changed.

What changed — and why it matters

01
A 429 is not an error like the others

A rate limit isn't a bug — it's capacity you're paying for and not getting. I split it from real failures: amber, not red. An engineer logs both as 4xx/5xx; knowing they're triaged differently is the judgment.

02
Tokens lead because cost leads

This log answers two questions: what failed, and what's expensive. Token counts lead the row because the cost question comes back every week — the request ID is context, not the story. That's a call about which number gets read first.

03
The footer counts what you can't see

Nine visible rows can hide 18,000 requests of trouble. The footer totals failures and rate limits across the whole window — so a clean first page never lies to you. Deciding what the page must admit is the design.