DeepSeek released V4.1 Flash on September 10 with native image understanding and lower API prices. Existing users face a more immediate change: requests to V4 Pro will start reaching the new Flash model on September 14. Applications can therefore change behavior without a configuration change.
The release announcement describes Flash as the smallest member of a new architecture family. DeepSeek is positioning it as a replacement for its previous flagship, but the published results support a narrower conclusion: strong gains on agent tasks, with remaining weaknesses on some reasoning evaluations.
The API migration is already underway
The new API name is deepseek-flash. DeepSeek has retired V4 Flash and V4 Flash Vision Exp; their old names temporarily route to V4.1 Flash. V4 Pro follows at 04:00 UTC on September 14, 2026, when deepseek-v4-pro requests will receive V4.1 Flash at Flash prices until V4.1 Pro launches. DeepSeek has not supplied a date for that next release in this announcement.
That is a behavior change even for applications whose requests continue to succeed. Teams maintaining regression suites should record the migration boundary and rerun representative tasks. Check the outputs as well as whether requests succeed.
It also gives August’s V4 Pro release and pricing changes a short shelf life as deployment guidance. Those details remain part of the release history; the new routing rules determine what customers will actually receive.
What the new architecture changes
DeepSeek’s model card describes a 552B-backbone-parameter mixture-of-experts model with a one-million-token context window. Its causal encoder-decoder activates 8B parameters during input processing and 16B during generation. The card separately lists 196B parameters of Engram conditional memory, so the backbone figure should not be treated as a complete deployment-memory estimate.
DeepSeek reports a global KV cache of 890 bytes per token, roughly a quarter of V4 Flash’s footprint, and persistent cache storage around one eighth as large. These are cache reductions, not equivalent reductions in the memory needed for the whole model. Our KV-cache guide explains why that distinction matters for long conversations and concurrent requests.
The repository and weights carry an MIT license. That permits broad reuse, subject to its terms; it does not establish that an existing local inference stack can run the new architecture efficiently.
Cheaper tokens, with two price windows
The published dollar rates are per million tokens:
| Tokens | Off-peak | Peak |
|---|---|---|
| Cached input | $0.003 | $0.006 |
| Uncached input | $0.15 | $0.30 |
| Output | $0.60 | $1.20 |
Peak hours are 01:00–04:00 and 06:00–10:00 UTC, Monday through Friday. All other hours are off-peak. The new rates took effect at 04:00 UTC on September 10.
For a simple cost illustration, one million uncached input tokens plus one million output tokens costs $0.75 off-peak or $1.50 at peak rates. This is arithmetic using the posted rates, not a measured cost per completed job. Retries and longer generated answers can change the bill even when the token price falls.
Agent gains come with reasoning tradeoffs
DeepSeek reports 90.6 on Terminal-Bench 2.1 and 74.2 on DeepSWE v1.1. However, its model card also gives V4.1 Flash 90.9 on GPQA Diamond against V4 Pro’s 92.4. Flash therefore trails Pro on at least this reasoning evaluation.
These are vendor-reported results. The card uses maximum reasoning effort, and agent scores depend on the surrounding software. On DeepSWE v1.1, the reported result is 74.2 with mini-SWE and 65.6 with Codex. That difference makes the harness part of the comparison.
There is also a release-day documentation discrepancy: the changelog lists NL2Repo-Bench at 65.4, while the model card lists 64.0. Neither figure should become an unqualified headline until DeepSeek reconciles them.
For an existing DeepSeek API deployment, the useful next step is a focused regression run before September 14: the same documents, tool definitions, acceptance criteria, and spending measurements used for V4 Pro. Lower prices and native vision make Flash worth testing for multimodal agent work, with output quality measured alongside the bill.
Sources
- DeepSeek V4.1 Flash release announcement — DeepSeek
- September 10 changelog — DeepSeek
- Models and pricing — DeepSeek
- V4.1 Flash model card — DeepSeek
- Model repository license — DeepSeek