DeepSeek V4.1 Flash: 'Opus-class but cheaper than Haiku' — and a professor mulls reselling B300 tokens
First seen on X 18 hours ago@cneuralnetwork ♥ 5,980
Users testing China's open-weight DeepSeek V4.1 Flash-1017 keep saying it matches top models at a fraction of the cost. Researcher @AlicanKiraz0 said at max effort it is "on par with Opus 5.5 and GPT-6.1 Sol High, while being cheaper than Haiku or Luna." Another said it equals Opus 4.8, the market leader a few months ago; a developer reported burning 12M tokens for about $0.18.
The low price is spilling into odd uses: CMU professor @gneubig mused that instead of research, he might resell DeepSeek tokens from his new B300 machine to fund his lab. DeepSeek says it raised performance while keeping API prices flat, and will start per-request billing on its Search/Fetch APIs on Nov 1.
“DEEPSEEK IS SO CHEAP!!!! i burnt 12M tokens only for 16rs”
Reactions (2)
- @mathew_noel♥55
Which inference provider? Because I don't yet feel comfortable with them using my data. - @cneuralnetwork♥139
deepseek itself
