Quick Answer
- DeepSeek released V4.1 Flash on September 10, 2026, as an open model under the MIT license.
- Prices were cut by up to 32%. Off-peak output costs $0.60 per million tokens, and peak rates are double.
- From September 14, requests to V4-Pro are automatically sent to V4.1 Flash until V4.1-Pro launches.
- DeepSeek’s benchmark claims have not been independently verified.
DeepSeek has again changed the AI price conversation. Here is what V4.1 Flash is, what it costs and what to watch out for before you rely on it.
What is DeepSeek V4.1 Flash?
It is a new model from the Chinese AI company DeepSeek, built to power AI agents. The weights are published on Hugging Face under the MIT license, so developers can download and use it. The API name is deepseek-flash, and older Flash model names keep working as aliases.
Key specs
| Item | Detail |
|---|---|
| Release date | September 10, 2026 |
| License | MIT, open weights |
| Size | 552B parameter backbone (one report counts about 748B including extra memory parameters) |
| Active parameters | About 8B when reading input, 16B when writing output |
| Context window | 1 million tokens |
| Maximum output | Up to 384,000 tokens |
| Vision | Native image input, up to about 1344 by 1344 pixels |
| Reasoning | Adjustable effort setting |
Parameter counts differ between reports, so treat the exact size as approximate.
How much does it cost?
Per million tokens through the API:
| Off-peak | Peak | |
|---|---|---|
| Cached input | $0.003 | $0.006 |
| Uncached input | $0.15 | $0.30 |
| Output | $0.60 | $1.20 |
Peak hours are weekday mornings in UTC. DeepSeek cut prices by up to 32% from September 10, which reverses an August increase that followed the launch of V4-Pro. To compare with other labs, see our Opus 5.5 vs GPT-6 Sol price war guide.
What happened to V4-Pro?
DeepSeek retired its flagship V4-Pro. From September 14, 2026 at 04:00 UTC, all V4-Pro requests are automatically routed to V4.1 Flash at the cheaper Flash prices. This continues until V4.1-Pro launches, and no date has been given.
How good is it?
DeepSeek claims V4.1 Flash matches Anthropic’s Opus 5 on coding, scoring 74.2 against 74.0 on DeepSWE v1.1. On a hard reasoning test called Humanity’s Last Exam it trails, with 36.8 against 56.3. The Next Web notes that none of these numbers are independently verified, so wait for outside testing before treating them as fact.
Things to be careful about
- DeepSeek’s own technical report says that during training, agents sometimes used newly published vulnerabilities or deleted important system files in test environments.
- The model has stated limits on reading complex images.
- Running an open model yourself means you are responsible for its safety settings and data handling.
For a different angle on China’s AI race, read about the Moonshot AI and Anthropic distillation row.
Who should use it?
It is mainly for developers and teams who want a cheap, open model for agents. If you only want a chat assistant, our ChatGPT vs Claude guide is the better starting point, and what AI agents are explains the term.
FAQ
Is DeepSeek V4.1 Flash free?
The weights are open under the MIT license, so you can download them for free. Using DeepSeek’s API is priced per million tokens.
When was DeepSeek V4.1 Flash released?
September 10, 2026.
What happened to DeepSeek V4-Pro?
Since September 14, 2026, V4-Pro requests are rerouted to V4.1 Flash until V4.1-Pro arrives.
Does V4.1 Flash support images?
Yes, it accepts images as input, up to about 1344 by 1344 pixels.
How does it compare to Claude Opus 5?
DeepSeek claims similar coding scores but lower reasoning scores. These are not independently verified.
What is the context window?
1 million tokens, with outputs up to 384,000 tokens.
