DeepSeek released V4.1 Flash—the junior model in a new architectural lineup with native multimodality (text, images, and audio are processed through a single pipeline rather than separate adapters). At launch, pricing remained at beta-period levels: 0,22 dollars per million input tokens for a cache miss and 0,66 dollars per million output tokens, with a context window of 1 million tokens. This is cheaper and, at the same time, faster than the previous V4 Flash version. On 10 September 2026, the day of the official release, DeepSeek cut prices once again—to 0,15 dollars for input and 0,60 for output during off-peak hours, with prices doubling during peak hours (from 01:00 to 04:00 and from 06:00 to 10:00 UTC on weekdays). According to the company’s internal tests, V4.1 Flash has already surpassed V4 Pro in performance, cost, and speed—and after launch, all requests to V4 Pro are automatically routed to V4.1 Flash and billed at its rates.
How the model was tested
The model was not run on benchmarks (none yet existed for such a new version), but on practical vibe-coding tasks—that is, generating working code from a single detailed prompt with the highest level of “reasoning.” Four varied tasks were selected: a low-poly 3D ice cream object, an aquarium with fish, a browser-based operating system simulator, and a platformer inspired by Super Mario. Each time, the generation time and number of tokens used were recorded—these are the main figures to consider when evaluating a model at this price point.
3D ice cream: fast and nearly flawless
Generation took 3 minutes 32 seconds and 53 000 tokens. The result was a cone with chocolate glaze, a cherry, and drips running down in lines and forming drops, along with a proper highlight on the ice cream’s surface. The main flaw was that the inside of the top section was not rendered; the geometry remained unfinished there. For a model in this price segment, the result was rated as one of the best of all four cases.
Aquarium: from the first version to a refined version in 150 000 tokens
The first version of the aquarium was assembled in 2 minutes 25 seconds using 48 000 tokens: fish, rocks, plants, air bubbles, and rippling water, with no complaints about the basic scene. A follow-up prompt expanded the scene by adding a snail, crayfish, and crab, all with correctly implemented movement mechanics—the crab moves sideways, the crayfish can move backward and forward, rather than simply repeating the fish animation. Both iterations took a total of 8 minutes 55 seconds and around 150 000 tokens.
OS simulator: surprisingly many functional applications
The most feature-rich project took 114 000 tokens. The resulting web interface imitates a desktop OS with light and dark theme switching, a file manager (documents, images, and downloads) with image previews, a text editor, a fully functional calculator, a clock with a stopwatch and timer, notes that can be deleted, a working terminal (it accepts commands and displays the current date), and a music player that actually plays music and looks correct—according to the tester, this is rare even among more expensive models. A functional Minesweeper game was also built separately. Among the decorative elements, the weather widget was implemented as a placeholder without real data.
Platformer: playable, but rough around the edges
The Super Mario-style game took 4 minutes 52 seconds to prepare and cost 51 000 tokens. Coins can be collected with sound, enemies can be defeated, jumps work correctly, and the background includes clouds, a finish-line flagpole, and a castle at the end of the level. On the downside, some decorative background objects (flowers and rocks) are displayed incorrectly, as if they were floating in the air; the pipes are not detailed enough, and the character’s respawn point after death is set incorrectly.
What to use it for
- For visual prototypes and simple interactive scenes (3D objects, mini-games, and interface mockups), V4.1 Flash delivers an experience comparable to more expensive models—and at a price several times lower.
- The model is also suitable for targeted code edits and tasks without complex architecture: it handles non-critical changes and moderately complex fixes competently.
- It is still too early to rely on it as the primary model for architectural decisions and complex refactoring—the tester describes it as an auxiliary model rather than a replacement for the main tool used in their workflow.
- Before budgeting a project using the old figures ($0,22 / $0,66 per million tokens), check the current pricing: after 10 September 2026, DeepSeek introduced lower off-peak rates and a distinction between peak and off-peak hours.
Compare models before you start
The service sets its plans, limits and model catalog. If they differ from this article, contact us so we can update it and record a new review date.
Browse modelsAffiliate link: your price stays the same and the project earns a commission.