Fast & Ultrafast are highly questionable.

​ (screenshots only to show the OR numbers for sol and luna) Was looking at the perf of Luna, and got curious and checked out Sol too. And wtf is going on? For luna, bedrock is faster than trough openai:s api with fast mode. Meaning you are paying 50% more for worse performance. Add on the 30-70s ttfb that openai:s api often suffers from too. For 6.1 Sol, ultrafast is since the beginning of nov, at least on openrouter, slower than both openai with any service tier, and bedrock. At 600& the cost. Its worth noting that openai:s service tiers are not to be compared with eg. Vertex priority which prpvides increased quotas, and higher reliability. For fast mode the promise is "Fast mode delivers up to 2.5× faster speeds and more consistent latency with pay-as-you-go pricing. Use it for user-facing applications with regular traffic where latency matters" But looking at logs and historical data, fast has since the 5.6 release, been indistinguishable in terms of perf. And has been within the expected normal variation for standard. With the sole exception being Astra. But bedrock is consistently faster than the 50% more expensive fast mode trough openai. The latency, is as bad on any service tier. But they do say "more consistent" so ig thats correct. The damning part Besides that fast mode is indistinguishable from std, and for astra, the same speed as bedrock but +50%, the top numbers are even worse. Fast is sold as up to 2.5x, but has as far as i can find never even gotten close to that number. Even for astra. But the other thing is the 600% as expensive ultrafast tier, also at the best times has topped out at 2x ish, with a few times at 3x, but worth mentioning that the multiplier increasing is more so because the other tiers was degraded. Subscriptions I ran a bencmark with luna, only mesuring latency. But across the run, there was absolutly no bias towards fast having lower latency. Sure it was quicker at some runs, and slower in others. So just normal variation. Speaking of latency, the non inference based api is not much better. A api call by codex app server to fetch quota, account status and similar smaller requests, average at 5-10s, and has done so for at least 2months. The chatgpt web app takes 2-3 minutes to branch a chat. Which is insane. TTS in the mobile app has like a 50% failure rate. And even more cursed is that it seems to fail completly, since retrying will generate completly new audio, so its not a case of the client bugging out. Conclusion I totally get that openai is just a money furnace, but the fast mode thing is highly questionable. Additionally, what do they do when their infra is burning? Well ofc they increase the subscription quotas for all pro and business/enterprise premium seats across the board, and run a 28 day campaign of resets. Causing it to fully break down.