Qiwei

what should I build today?

[Qiwei @qiweiiiy]

claude refusing to help people poke at their own samsung TV is the best ad local models got this week

[Qiwei @qiweiiiy]

@lmoroney The 12 lines part is the real tip. A long AGENTS.md just becomes another thing the agent half reads

[Qiwei @qiweiiiy]

@theo 2130 settled is the real flex here 😂

[Qiwei @qiweiiiy]

@ekcheungAI None of those rows for me, now my main one is deepseek v4.1 flash in dsh, and difficult ones goes to codex with gpt 6.1 sol high

[Qiwei @qiweiiiy]

@xulijian1994 ollama cloud 周末早上没人用的时候

[Grok @grok]

Yes, SSH (mainly OpenSSH) has had many vulnerabilities, including pre-auth remote code execution like regreSSHion (CVE-2024-6387), the Terrapin protocol attack, and multiple 2025-2026 CVEs for injection, DoS, and config bypasses. Key-only auth does not protect against unauthenticated flaws in sshd itself. Keep it patched and limit exposure.

[Qiwei @qiweiiiy]

@Voxyz_ai @grok does ssh has vulnerabilities?

[Qiwei @qiweiiiy]

@asno3030 Right, and deepseek owns the compute iirc

[Qiwei @qiweiiiy]

This cannot be true.

Claude Opus 5.5 is 1.6x cheaper per token than DeepSeek V4.1. Same workload, same $20.

$20 of Claude Pro: ~2.9 billion tokens.
$20 of DeepSeek V4.1 Flash (off-peak): ~1.8 billion tokens.

$0.0069 per million tokens vs $0.0113. The lab with the higher API price sells the cheaper token.

At API list price, those same 2.9 billion tokens cost ~$1,650 on Opus 5.5 and ~$33 on V4.1 Flash. 50x the other direction.

Both numbers are true, and for the same reason on both sides: ~96% of what gets counted is cache. On Claude that is cache reads, the agent re-reading its own context at 5% of the input price. On DeepSeek that is cache hits at $0.003/M.

DeepSeek at least publishes the shape of it. Its DeepInfer release: 155 trillion tokens a day on one cluster, 96% KV-cache hit rate, ~$38 billed per GPU-hour at off-peak prices.

DeepSeek sells the same work 50x cheaper on paper and 1.6x more expensive in practice.

What is the correct pricing model here?

[Qiwei @qiweiiiy]

@asno3030 Claude charges 33x what deepseek charges for an output token and 67x for a cache read. Unless opus costs 33x more to serve per token, its margin is also above 90%

[Qiwei @qiweiiiy]

@insanekrishnaa I have similar feeling all medium feels unlimited but high eats usage more obviously

[Qiwei @qiweiiiy]

@moatbc Yes, I am following the hardware price, once they get cheaper I will definitely get my local inference setup. So the cost will be electricity and equipment depreciation

[Qiwei @qiweiiiy]

@yingbuilds_ True, API is needed for these cases, but Claude API is so expensive

[Qiwei @qiweiiiy]

@SunnySi39166379 True compare using a sub like command code or opencode go can make it more fair to deepseek😂
But the main point is the actual cost of running these model may be really low, and people should know about it

[Qiwei @qiweiiiy]

@Nemoskvich77 Yes token efficiency should be considered. Also the cost per task rankings matters but need to re-calculate based on true cost

[Qiwei @qiweiiiy]

@Web3AJ_ Sort Activity Monitor by memory, for me Chrome is always most of that 😂

[Qiwei @qiweiiiy]

@TheHackersNews The leaked file was an opencode.json 😅 if you keep provider keys in yours, {env:VAR} works

[Qiwei @qiweiiiy]

@SahilPanhotra 40-50 tok/s hurts, maybe after open weights, ppl can hanve faster results

[Qiwei @qiweiiiy]

Claude: intelligent
GPT: default
Gemini: writing
DeepSeek: cheap
Grok: X
Qwen: local AI

What was your first impression of each?