DeepSeek V4.1 Flash
A very cheap, very efficient open-weights model from China, and plenty for a personal assistant agent.
Quality
Modality
multimodal
Context
1M tokens
Access
open weights
Fabian's Take
"This is the model behind my OpenClaw assistant. It's very efficient and does everything I need from a personal assistant agent. I find it hard to tell the difference from the previous version in day-to-day use."
DeepSeek V4.1 Flash replaced V4 Flash on 10 September 2026; the current snapshot is from 17 October. If you still call the old model name through the API, your requests go to V4.1 Flash anyway.
Why it’s interesting
The price. At peak hours, a million output tokens cost $1.20; outside China’s working hours (DeepSeek lists the exact peak windows) it’s half that, and repeated input that hits the cache costs almost nothing. That makes agents that run around the clock, or jobs you repeat hundreds of times, affordable in a way the frontier models aren’t.
It’s also more capable than the price suggests. By DeepSeek’s own benchmarks it trails the frontier models on the hardest reasoning tests and keeps up on some coding ones. For what I need from a personal assistant agent, I haven’t hit its limits.
Open weights, but not for your laptop
The weights are published under the MIT license, so companies can host it themselves. That’s a server job: the model has well over 500 billion parameters. If you want something that runs on your own machine, look at Gemma 4 instead.
The Verdict
Best for: Agents that run all day, and any task you need to run hundreds of times without watching the bill.
Pros
- Very cheap: output costs $1.20 per million tokens at peak and half that off-peak
- Efficient enough to run a personal assistant agent all day
- Open weights under the MIT license
- Now reads images as well as text
Cons
- Prices depend on the time of day, which makes costs harder to predict
- Behind the frontier models on the hardest reasoning tests, by DeepSeek's own numbers
- Too large to run on your own computer, despite the open weights
- DeepSeek's own API stores data in China; for anything personal or work-related, use a provider that hosts the open weights elsewhere, or check where your requests go
Specs
- Pricing $0.30/M input, $1.20/M output at peak · half price off-peak ($0.15/$0.60) · cache hits $0.003–0.006/M
- Cost Tier budget
- ⚡Speed Tier fast
- License MIT
