xiaomi mimo referral code for free API tokens: LGWZCN
v2.5: 0.14 input, 0.0028 cache input, 0.28 output (~dsv4-flash)
v2.5-pro: 0.435, 0.0036, 0.87 (~dsv4-pro)
im still trying it myself, but its as cheap as DSV4 now, and from benchmarks seems a bit stronger
@HelloVyom America’s biggest export is the American dream, and there is no better place to realise that than the USA. China’s fundamental belief system is different, and advancing their own national goals usurps capitalistic individualism
@deadrosesxyz@aviggiano Imo, contests should not have submission cost, but platform takes a slightly larger fee and require PoC. It’s operationally more intensive, but restricting the onboarding pathway is not a good tradeoff otherwise. Also, lower false negative rate is worth higher false positive rate
@W98AB@shguke@gnievchenko I’m not doubting you, maybe it’s a Claude issue, or you’ve just been unlucky? I also can’t know exactly how you use them, I can only share what works for me. I apologise if you thought that was me calling you retarded or whatever
@W98AB@shguke@gnievchenko Sonnet is expensive, deepseek v4 flash costs .14 for input per million, and .28 output per million. Cached input is basically free it’s so cheap. Straight forward implementation stuff is easy, but not necessarily one-shot. Feed it back errors and correctness issues and it��ll fix
@W98AB@shguke@gnievchenko You could also do this for free using Gemma 4 31B via a free google ai studio account, but it’s a bit slower and may take more iterations
@W98AB@shguke@gnievchenko Implementation with clear instructions is easy, but even the best models may not one-shot. Try something like deepseek V4 flash (off/high), and instruct it to verify correctness and fix errors. It’ll cost <~$0.10. Save stronger models for deep reasoning/planning on hard problems
deepseek v4 is very good for the price
i ran out of codex so now im using roo code with deepseek v4 flash and ive used about 50m tokens today, and spent around $0.50 (a lot of cache hit input)
it's no gpt 5.5, but it fills the void until my usage resets
@0xngmi I agree broadly, but that’s not entirely addressable. Some things are inherently qualitative and subjective (eg see how solana, ethereum and Bitcoin maxis discuss decentralization), and attempting to quantify them can lead to a Goodhart’s law issue