At a glance
- The September 10 announcement introduces V4.1 Flash and the deepseek-flash ID.
- Native vision and architecture improvements are vendor-reported capabilities.
- Pricing distinguishes cache hits, misses, output and time periods.
What changed
DeepSeek announced V4.1 Flash on September 10, 2026, describing architectural efficiency, native vision and revised pricing. Its recommended API identifier is deepseek-flash, which the current pricing table also maps to V4.1 Flash. Source 1 Source 2
The announcement describes a 552B-parameter mixture-of-experts architecture with different activation sizes for input and output. These are vendor-reported design and performance details. Parameter counts alone do not establish accuracy or cost for a particular workload, and no independent benchmark was performed for this article.
Evaluate vision with known answers
The pricing page lists vision support, a 1M-token context and up to 384K output tokens for Flash. Context and output ceilings still require a valid combined request budget. Source 2
Try charts, error screenshots or interface documentation. Ask for visible facts before conclusions, and have the model mark unreadable regions. Include tables with known row and column values so that subtle recognition errors become visible.
Check the client’s image payload and the provider route as well. A failed request can reflect integration rather than the model’s visual reasoning.
Read each pricing dimension
Official US-dollar rates per million tokens at verification were as follows. These are not Youjie channel quotations. Source 2
| Item | Peak | Off-peak |
|---|---|---|
| Input cache hit | $0.006 | $0.003 |
| Input cache miss | $0.30 | $0.15 |
| Output | $1.20 | $0.60 |
The page defines peak hours as Monday–Friday, 01:00–04:00 and 06:00–10:00 UTC, excluding Chinese public holidays. That corresponds to 09:00–12:00 and 14:00–18:00 in Beijing. Confirm actual billing rules for boundary cases.
A budget example
For one million uncached input tokens and 200,000 output tokens, peak cost is 1 × $0.30 + 0.2 × $1.20 = $0.54. Off-peak cost is 1 × $0.15 + 0.2 × $0.60 = $0.27, excluding retries, tools and channel charges.
Deferrable batch work may benefit from scheduling. Real-time services still need to meet response requirements, and cache behavior should be measured from usage records.
Keep capability and migration decisions separate
Historical Flash aliases and Pro routing require additional checks because official materials differ. The companion migration article explains that uncertainty.
Record accuracy, latency, cache hits and the cost of completed tasks before adopting the new version. Architectural claims are useful context, but they are not a substitute for workload evidence.
Sources & verification
Primary materials consulted for this article. Vendor measurements are not independent Youjie tests.
- DeepSeek-V4.1-Flash: Smarter, Faster, More Efficient ↗DeepSeek · Source date 2026-09-10 · Checked Oct 16, 2026
- DeepSeek Models & Pricing ↗DeepSeek · Documentation · Checked Oct 16, 2026