DeepSeek V4.1 Flash is now the company’s default API model, released on 10 September 2026 with open weights, a 1-million-token context window and lower prices. Its change log says DeepSeek V4 Pro will keep running after 14 September with billing unchanged, contradicting reports of a forced reroute. No V4.1 Pro date has been given.
DeepSeek V4 Pro Keeps Running After 14 September
Teams calling deepseek-v4-pro in production do not have a 14 September deadline, according to DeepSeek’s own documentation. The change log entry for the V4.1 Flash release states that the company has “decided to continue providing API services for DeepSeek V4 Pro after September 14, 2026, with the billing method remaining unchanged”.
That matters because several developer write-ups published between 9 and 11 September said the opposite. Digital Applied and the AI support vendor eesel both reported that from 04:00 UTC on 14 September, equivalent to noon Beijing time, every deepseek-v4-pro request would be served by V4.1 Flash and billed at Flash prices until a V4.1 Pro model arrived. As of 12 September 2026, DeepSeek’s change log describes no such rerouting for the Pro endpoint.
DeepSeek, the Hangzhou-based AI company whose low-cost models drew global attention in early 2025, has not published a statement explaining the difference, and the reroute plan may simply have been dropped after those pieces were written. Developers relying on Pro-level output should treat DeepSeek’s change log and pricing pages as the source of record rather than secondary summaries.
What V4.1 Flash Is
It is the smallest model in DeepSeek’s new architecture family, and unusually for a “small” model, it is large on paper. The published model card gives these specifications:
- Parameters: 552 billion in the backbone, with 8 billion active per token during prefill and 16 billion during decode.
- Architecture: a Causal Encoder-Decoder design, described as a 40-layer Transformer built as a 20-layer causal encoder followed by a 20-layer decoder.
- Context: up to 1 million tokens in, and up to 384,000 tokens out.
- Inputs: native multimodal visual understanding, so images and text go in and text comes out.
- Licence: MIT, with weights published on Hugging Face.
- API name: deepseek-flash.
On the model card’s own scores, V4.1 Flash reaches 74.1 on MMLU-Pro, 79.4 Pass@1 on HumanEval and 74.2 resolved on DeepSWE v1.1. DeepSeek says it retired the previous generation because testing showed the new model had surpassed it on performance, cost and speed.
What It Costs to Call
DeepSeek charges two rates depending on the hour, and the cheaper one is the default outside business hours in China. Peak hours run 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday to Friday; off-peak prices are half the peak prices. Figures below are per million tokens, from DeepSeek’s published price list as of 12 September 2026.
| Charge | deepseek-flash (off-peak / peak) | deepseek-v4-pro (off-peak / peak) |
|---|---|---|
| Input, cache hit | $0.003 / $0.006 | $0.022 / $0.044 |
| Input, cache miss | $0.15 / $0.30 | $0.66 / $1.32 |
| Output | $0.60 / $1.20 | $1.98 / $3.96 |
On cache-miss input, Flash is about 4.4 times cheaper than V4 Pro; on output, about 3.3 times cheaper. Prompt caching, which reuses tokens the model has already processed, is where the gap is widest at roughly seven times.
Model Names That Have Changed
Two older endpoints are gone, and DeepSeek is catching their traffic for now rather than failing it.
- deepseek-v4-flash: retired, temporarily routed to V4.1 Flash.
- deepseek-v4-flash-vision-exp: retired, temporarily routed to V4.1 Flash.
- deepseek-flash: the production name to migrate to.
- deepseek-v4-pro: unchanged, and billed at Pro rates.
Because the routing of the retired names is described as temporary, code still pointing at them needs updating even though it works today.
What DeepSeek Has Not Said
There is no announced release date for V4.1 Pro, the larger model the change log implies is coming. DeepSeek has not said how long the retired Flash names will keep resolving, has not commented on the conflicting reports about the Pro endpoint, and has not published throughput or rate-limit figures alongside the new prices. The company’s low-cost pricing has been a competitive lever since the release of the model that unsettled the AI industry in January 2025, and it has given no indication of how long these rates hold.
Frequently Asked Questions
Common questions about DeepSeek V4.1 Flash and the V4 Pro endpoint.
Is DeepSeek V4 Pro Being Shut Down on 14 September 2026?
No, according to DeepSeek’s change log, which says API services for DeepSeek V4 Pro continue after 14 September 2026 with the billing method unchanged. Several developer write-ups published 9 to 11 September reported a reroute to V4.1 Flash instead.
What Does DeepSeek V4.1 Flash Cost per Million Tokens?
Off-peak, $0.15 for cache-miss input, $0.003 for cache-hit input and $0.60 for output. Peak rates are double: $0.30, $0.006 and $1.20. Peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday to Friday.
What Is the Context Window of DeepSeek V4.1 Flash?
One million tokens of input, with a maximum output of 384,000 tokens, according to DeepSeek’s pricing documentation.
Are the DeepSeek V4.1 Flash Weights Open?
Yes. The model is published under the MIT licence with weights on Hugging Face, so it can be downloaded and run outside DeepSeek’s API.
Which API Model Name Should I Use Now?
Use deepseek-flash. The older deepseek-v4-flash and deepseek-v4-flash-vision-exp names are retired and only temporarily routed to V4.1 Flash, so code pointing at them should be updated.
