Maharba

[Maharba @maharbaisc]

@Gosleepriya 949

[Maharba @maharbaisc]

@ChialFelix Excelente, me despertó antes de que llegara la réplica de 6.0 en la madrugada

[Maharba @maharbaisc]

@traficocpanama Están diseñados así precisamente para que no colapsen

[Maharba @maharbaisc]

@bijanbowen Fun fact, you can get fable performance and at Haiku speed and prices if you set Fable as the orchestrator and spawn as many Haiku sub agents as fable can manage. This should be the way new models are tested

[Maharba @maharbaisc]

@perma_eevee @ChialFelix Es una culebra, viven debajo de la tierra, comen lombrices y otros insectos.

[Maharba @maharbaisc]

@bridgemindai Opus 5.5 is INSANE

[Maharba @maharbaisc]

@bridgemindai @bridgebench That would be on mid effort on high it burns tokens just as fast as Astra

[Maharba @maharbaisc]

@elcuaracom Porecito 😭😭😭😭😭 - Tospa

[Maharba @maharbaisc]

@ziwenxu_ This is what you should be striving for, My 20X plan is about 80% after reset.

[Maharba @maharbaisc]

@ziwenxu_ I've spend about 1.5T tokens on Deepseek per 20 millions or so in Astra, that's the real efficiency metric.

[Maharba @maharbaisc]

@bridgemindai I've reduced my spend and speedup my dev cycle using this formula Astra Max to diagnose and create a plan, Astra light to orchestrate a swarm of Deepseek 4.1 Flash agents, 12 hours of work are roughtly equivalent to 8% usage with that setup, working on a c++ app.

[Maharba @maharbaisc]

@DesignArena Its becoming incresingly clear that for agentic work you need Astra, for everyday coding needs DeepSeek is sufficient, I've been using it for a couple of days and have not missed Astra a bit on my C++ project.

[Maharba @maharbaisc]

@bridgemindai Indeed, same here!

[Maharba @maharbaisc]

@elshayib_ I've found that the best way to use it is under human supervision. It likes to go on tangents exploring and even coming up with unlikely extreme edge cases and spend hours trying to solve them. If you give it a direct, concrete task, it does it very well and very fast and cheap.

[Maharba @maharbaisc]

@arena @OpenAI Still amazed Qwen3.8 - 27B is trading blows with trillion parameters models (may be trillion dollars too).

[Maharba @maharbaisc]

@FaztTech Recuerdan DxOMark? que hacia ratings de camaras para los celulares? AA es eso, para IA

[Maharba @maharbaisc]

@DanDr1s AA is the DxOMark of the AI world

[Maharba @maharbaisc]

@IvanMon07722658 No es de extrañar que estes del lado de los abogados del narco

[Maharba @maharbaisc]

@BrenBuilds @OpenAI These are all current generation models with different reasoning levels, this is not evidence of a wall, all of the contrary