Image: Intelligent Living / intelligentliving.co DeepSeek Launches V4.1 Flash and Will Route All V4 Pro Traffic to It Starting September 14
DeepSeek's new V4.1 Flash model beats V4 Pro on benchmarks while running at Flash pricing. Starting September 14, V4 Pro API requests automatically redirect to it.
DeepSeek quietly launched V4.1 Flash today, its most capable model yet, and announced that starting September 14 at noon Beijing time, all V4 Pro API requests will automatically route to it at V4.1 Flash pricing.
The auto-routing decision is the bigger news. If you’re paying for V4 Pro today, you’ll get a better model and pay less for it in four days without changing a line of code. Legacy endpoint names deepseek-v4-flash and deepseek-v4-flash-vision-exp will temporarily redirect to V4.1 Flash as well.
What Changed in the Model
V4.1 Flash uses a new architecture: a 552-billion parameter backbone combined with 196 billion Engram parameters, with only 8 to 16 billion parameters active per token during inference. That mixture keeps costs down while letting the model draw on a much larger knowledge base.
Native multimodal visual understanding is built directly into the base architecture, not bolted on. The context window is 1 million tokens.
Benchmarks tell a clear story. V4.1 Flash hits 90.9 on GPQA Diamond, a Codeforces rating of 3471, and 90.6 on Terminal-Bench 2.1. It beats V4 Flash, V4 Flash Vision Exp, and V4 Pro on every shared benchmark DeepSeek tested.
A two-day beta endpoint called deepseek-v4.1-flash-expires-on-0910 ran September 8 through 10. The expiry date in its name turned out to be the production launch date.
Pricing
The price cuts range from 11% to 57% depending on token type. Cached input prices dropped roughly 60% compared to V4 Flash. DeepSeek hasn’t published dollar-denominated prices in today’s announcement, but the yuan pricing moves put V4.1 Flash well below what V4 Pro cost before.
Use the endpoint deepseek-flash to always call the latest Flash-tier model going forward.
Who Should Care
Any developer using V4 Pro will see the upgrade automatically on September 14. If you want to test V4.1 Flash before then, the model is live now via deepseek-v4.1-flash in the API.
For AI coding tool builders, the Terminal-Bench score of 90.6 puts V4.1 Flash in a very competitive position for agentic code tasks. For reference, Claude Fable 5.1 scores 55.8 on Terminal-Bench 4.0 and Mythos 5.1 reaches 60.9 — though those use a different version of the benchmark, so direct comparison isn’t clean.
DeepSeek is publishing the full technical writeup at api-docs.deepseek.com/updates.