Kale Nine

I tell computers what to do

[Kale Nine @nullmasterb]

@scaling01 A genuine question. In Artificial analysis Sonnet 5.5 uses 7 times more tokens than Astra. How is that possible? What is it doing?

[Kale Nine @nullmasterb]

@AC_Creative08 @count_of23995 @bridgemindai Stop baiting me with stupid posts. It was obvious that he just wanted to spend all the usage on stupid background tasks with many agents running concurrently. Auditing your whole codebase is a background task. Ultrafast is for more interactive use where the 6x price is worth it

[Kale Nine @nullmasterb]

@count_of23995 @bridgemindai I watched the stream. He ran a large amount of Astra agents in parallel, auditing all his codebases.

[Kale Nine @nullmasterb]

@bridgemindai Obvious bot...

[Kale Nine @nullmasterb]

@savvyntsev @robsonsmelo @albertgao @theo The 1.7 number actaully originates from their own support side numbers.

https://t.co/6e2I6MuzDS

480/280 roughly 1.7

[Kale Nine @nullmasterb]

@gjhyj_ @SimonasLTU1 I never use max. IMHO max is just for wasting time and tokens. I use medium for most tasks and xhigh for planning.

[Kale Nine @nullmasterb]

@itsvlady Sometimes I just want the fastest and cheapest answers for low intelligence and low importance interactive tasks

[Kale Nine @nullmasterb]

@gjhyj_ @SimonasLTU1 In benchmarks the tasks are deliberately hard and harnesses take many turns to complete the task. Lesser models take more turns to complete the task, increasing their cost. In daily agentic use, if the task is simple and both answer in one shot = cheaper token gives more usage.

[Kale Nine @nullmasterb]

@gjhyj_ @SimonasLTU1 I honestly I have the same expectation. It would seem logical that all the Anthropic 5.5 models would beat GPT6 models in "intelligence." As to cost per task I am not sure. OpenAI did the 50% discount and Anthropic models usually use more than 2x tokens (thinking + "tokenizer")

[Kale Nine @nullmasterb]

@miracache I gave Opus 5.5 a product idea and a task to choose optimal tech stack for it. It considered python, TS and Kotlin backend. It chose Kotlin and implemented with. I wish I had not YOLO-d that part 😅

[Kale Nine @nullmasterb]

@gjhyj_ @SimonasLTU1 Well to me it seemed to be an Anthropic vs OpenAi bait post and he used his Luna vs Opus example to "support" his "point" which is low...
It is fine that Luna loses to Mimo with similar API pricing, but via subscriptions you get up to 50 times lower pricing than API

[Kale Nine @nullmasterb]

@SimonasLTU1 I read and still do not understand why you compared Luna result with Opus 5.5 result. Please do explain.

[Kale Nine @nullmasterb]

@nikitathakur21 DB has index usage stats recording and actually usually the more expensives writes are not the main concern. It is the increased memory usage.

[Kale Nine @nullmasterb]

@ivanfioravanti I think closed more efficient models, specialized hardware and higher average utilization rates will keep cloud cheaper until local AI does not require any additional hardware.

[Kale Nine @nullmasterb]

@ivanfioravanti Based on what I have seen, GPT-6 luna max on oAI subscription is practically "free" compared to anything local. But I actually I do see the value in local models, where data must not leave premise or systems are air-gapped.

[Kale Nine @nullmasterb]

@anjanybuilds Jesus. All the steps are still the same, but AI does them and is better at it then humans are...

[Kale Nine @nullmasterb]

@ivanfioravanti And don't get me started on the hardware and electricity ROI.

[Kale Nine @nullmasterb]

@ivanfioravanti I am a software engineer. When you consider the latency, output speed, my work time needed to maintain and use it and value the hourly wage as does my employer, then it is stupid useless.