i ran a similar DeepSeek V4.1 Flash 1017 comparison myself after seeing this. 4 things i noticed: 1. the official API was consistently faster for me. OpenCode Go is cheap, but once the task gets longer the speed difference becomes hard to ignore. 2. cache efficiency was noticeably better direct. i was seeing 99%+ cached tokens much more consistently, while OpenCode was closer to the 92–96% range. 3. OpenCode also gave me more timeouts and wasted runs. that matters a lot more than the advertised 8x inference once i’m doing real work. 4. the biggest difference was iteration speed. with the direct API i can often make several attempts while OpenCode is still finishing the first one. i still think OpenCode Go is great value. but for serious work, i prefer DeepSeek direct.

i'm a founder building my entire GTM system on top of my product codebase, and it's constantly falling apart. who should i talk to? a GTM system has to hold: -> open-ended research, like figuring out who's on Lovable's team -> a pile of skills for text and video that need to work together -> analytics results and the interfaces to query them -> CLI tools and .env files everywhere what i'm struggling with: -> reliably finding research i've already done when a new task needs it -> getting the right skills to work together with the right context -> keeping analytics results connected to the tools that produced them -> changing the setup without losing track of how everything fits together Codex feels like the start of a golden age for GTM. i'm still struggling to build infra that's easy to change and has reliable context retrieval. who's built something like this that they actually use every day? i'd love to see how you've structured it, especially how you organize knowledge and connect skills. a repo or architecture walkthrough would help a lot. who should i talk to?

Today I hit 5,000 followers!! 🥳 5 days ago, I was at 4,500. I told myself: “Let’s try to hit 5K in the next month.” It took 3 days 😅 Still kind of crazy to me. Thank you to everyone who followed, replied, shared, or just kept showing up in my feed 🫶 Turns out growth can feel painfully slow… right until it doesn’t. On to the next milestone 💜

One Jev use case with HUGE value 100 people apply for one open role. Normally someone has to review every profile manually. Jev can classify all 100 almost instantly. In our benchmark, it finished in just 0.827 seconds. That was 4.26x faster than the next fastest option. It was also 2.6x cheaper in that test.

okay, quietly admitting this: we’ve spent a ridiculous amount of time thinking about how agents actually become part of a team. sandboxes matter. but agents can’t live in a sandbox forever. they need channels where humans + other agents are already working together. they need clear task flows — when to keep going, when to interrupt, when to ask a human for approval. all of this together becomes context. and increasingly, i think context is what turns an agent from a tool you use into a teammate you work with. so… why not try offloop

here’s the list of western companies moving ai workloads to chinese models: 1. lindy → deepseek v4 2. cursor → kimi k2.5 3. coinbase → glm-5.2 + kimi 2.7 4. shopify → qwen 5. airbnb → qwen 6. uber eats → qwen2 7. siemens → deepseek + qwen 8. chapsvision → qwen 9. microsoft → testing deepseek v4 it’s becoming a procurement story!!!