DeepSeek AI Ships V4.1-Flash for Real: New Architecture, Native Vision, and Peak Pricing That Is Finally Live
By the Deep Seek AI editorial desk · September 18, 2026 · 7 min read
Ten days ago the model existed only as an ID with an expiry date baked into it. On September 9, 2026 DeepSeek published the release notes: V4.1-Flash is a general availability model, it answers to deepseek-flash, and it comes with the first two-tier rate card DeepSeek has run on the V4 line.
The 60-second version
- DeepSeek-V4.1-Flash is GA, no longer a beta. API name:
deepseek-flash. - 552B-parameter Mixture-of-Experts with a "Causal Encoder–Decoder" split — roughly 8B active parameters on input, 16B on output.
- Native image understanding inside the model, so the separate vision variant is gone.
- V4-Flash and V4-Flash-Vision-Exp are retired. Their model names still resolve, but requests are served by V4.1-Flash.
- V4-Pro is winding down: since September 14, 04:00 UTC,
deepseek-v4-protraffic routes to V4.1-Flash at V4.1 rates until V4.1-Pro arrives — unless you ask DeepSeek to keep V4-Pro service running. - Peak and off-peak pricing is live from September 10, 04:00 UTC. Off-peak is half of peak.
- Weights and a technical report are on Hugging Face, as with previous DeepSeek releases.
What actually changed in the architecture
The headline number — 552B total parameters — matters less than how few of them run per token. DeepSeek describes V4.1-Flash as a causal encoder–decoder MoE: the encoder path activates on the order of 8B parameters to read input, the decoder path around 16B to write output. That asymmetry is the point. Prompt-heavy workloads — long documents, retrieval context, codebases — are the cheap half of the model to run, and the expensive half only engages for the tokens you actually generate.
The second structural change is memory. DeepSeek reports the KV cache is 4× smaller in HBM and 8× smaller on SSD than the previous Flash generation. For anyone self-hosting the open weights that is the number to plan against: it decides how many concurrent long-context sessions fit on a given card before you start paging.
Vision is now native. Previously, image input on the Flash line meant a separate experimental model, deepseek-v4-flash-vision-exp. That model is retired because the capability moved inside. DeepSeek reports benchmark results above V4-Pro — its previous frontier tier — which is why Pro is being routed away rather than kept as the premium option.
The new rate card, in full
Prices below are per 1,000,000 tokens in USD, taken from DeepSeek's official pricing page and last verified by a human editor on 2026-09-18. Off-peak is the tier that applies most hours of the week, so we quote it first.
| Model | Tier | Input (miss) | Input (hit) | Output |
|---|---|---|---|---|
| DeepSeek-V4.1-Flash | Off-peak | $0.15 | $0.003 | $0.6 |
| DeepSeek-V4.1-Flash | Peak | $0.3 | $0.006 | $1.2 |
| DeepSeek-V4-Pro | Off-peak | $0.66 | $0.022 | $1.98 |
| DeepSeek-V4-Pro | Peak | $1.32 | $0.044 | $3.96 |
Peak hours are 01:00–04:00 and 06:00–10:00 UTC, Monday to Friday — 09:00–12:00 and 14:00–18:00 Beijing time. Every other hour is off-peak, weekends included. That is 35 of 168 hours, so a workload spread evenly across the week pays roughly 21% above the off-peak rate; batch work scheduled outside the windows pays nothing extra.
A correction we owe readers. Until this release our pricing page described the peak-hour surcharge as announced but not active, because the official rate card showed one flat tier. That is no longer true: it went live with this release, and the surcharge is a straight 100% — off-peak is exactly half of peak. Every page on this site that quotes DeepSeek rates now reads from one verified configuration, so the change propagated everywhere at once.
What this costs in practice
A concrete example on V4.1-Flash off-peak: 1M input tokens on a cache miss plus 1M output tokens costs $0.75. Add a stable system prompt so 90% of input hits the cache and it falls to about $0.62. Run the identical call inside a peak window and it is $1.50. Caching and scheduling are now worth more than model choice for most bills.
Our DeepSeek pricing page carries the full table, the price history and a calculator with a peak-hour toggle, and the homepage comparator puts the same workload next to GPT, Claude and Gemini rates.
Migration checklist
- Switch model strings to
deepseek-flash. The old Flash names still resolve, but relying on a retired alias is technical debt with a deadline. - Drop
deepseek-v4-flash-vision-expand send images todeepseek-flashdirectly. - If you are on
deepseek-v4-pro, decide deliberately: accept the routing to V4.1-Flash, or contact DeepSeek to keep V4-Pro service and its billing. - Re-check your cost model against the two-tier card, and move batch jobs out of the two peak windows.
- Re-run your own evaluations. Benchmarks reported by a model vendor are a starting point, not a verdict on your workload.
Frequently asked questions
What is the API name for DeepSeek-V4.1-Flash?
deepseek-flash. The base URL stays https://api.deepseek.com. The expiring beta ID deepseek-v4.1-flash-expires-on-0910 no longer serves traffic. The older names deepseek-v4-flash and deepseek-v4-flash-vision-exp are still accepted but those models are retired, so their requests are served by V4.1-Flash.
How much does DeepSeek-V4.1-Flash cost?
Per 1M tokens in USD, off-peak: $0.15 input on a cache miss, $0.003 on a cache hit, $0.6 output. Peak rates are exactly double: $0.3, $0.006 and $1.2. Peak hours are Monday to Friday, 01:00-04:00 and 06:00-10:00 UTC; every other hour, including weekends, is off-peak. Rates last verified 2026-09-18.
Is the DeepSeek peak-hour surcharge finally active?
Yes. From September 10, 2026 at 04:00 UTC the official rate card carries two tiers, with off-peak priced at 50% of peak. Until then DeepSeek billed a single flat rate, and this site reported the surcharge as announced-but-not-active, which was correct at the time.
What happens to DeepSeek-V4-Pro?
V4-Pro is being phased out. Since September 14, 2026 at 04:00 UTC, deepseek-v4-pro requests route to V4.1-Flash and are billed at V4.1 rates until V4.1-Pro ships. Customers who ask DeepSeek to keep V4-Pro API service can continue on it with unchanged billing.
Does V4.1-Flash handle images?
Yes. Image understanding is built into the model rather than bolted on through a separate vision variant, which is why DeepSeek-V4-Flash-Vision-Exp is retired alongside V4-Flash.
Are the V4.1-Flash weights open?
DeepSeek published the model on Hugging Face together with a technical report, in line with its previous releases. Licence terms are stated on the model card, not here — check it before commercial deployment.