DeepSeek shipped V4.1 Flash on September 10, 2026 as a new 552B MoE / 1M-context architecture family (call deepseek-flash), not a cosmetic .1 bump. Flash pricing already applies. The Sept 14 Pro policy is disputed across official pages: the Sep 10 news post still announces deepseek-v4-pro remaps to Flash at Flash rates from 04:00 UTC / 12:00 Beijing on Sep 14, while live Models & Pricing, Change Log, and First API Call footnotes say V4 Pro continues after that date with billing unchanged. Re-verify flag (Sep 14 morning): before and after that clock, re-check those four live pages - flip-day authority is still OPEN as of Sep 13.
In short
- Shipped Sep 10, 2026 - new architecture family, not a point release
- Canonical ID:
deepseek-flash; legacy Flash / vision-exp aliases temporarily route to V4.1 Flash - 552B MoE; 8B prefill / 16B decode active; 1M context; max output 384K; vision on Flash
- Flash rates live since 04:00 UTC Sep 10; peak / off-peak schedule published
- OPEN CONFLICT on
deepseek-v4-proafter Sep 14 - news260910 remap vs live ops "Pro continues" (confirmed-vs-reported: both first-party, unresolved) - Sep 14 morning re-verify required on pricing, First API Call, Change Log, and news260910
- Checklist: re-read live docs - decide Flash default vs keep Pro path
- Not a frontier bake-off vs other labs' flagship models
Why isn't V4.1 Flash just a point release?
DeepSeek frames V4.1 Flash as the smallest model in a new architecture family with native visual understanding. That is a base-model story, even though the marketing label still says .1.
Vendor-published facts from the news post and the Hugging Face model card:
| Spec | Published value |
|---|---|
| Backbone | 552B-parameter MoE |
| Active compute | Causal Encoder-Decoder: 8B active for input/prefill, 16B for output/decode |
| Context | Up to 1M tokens |
| API max output | 384K |
| Vision | Native on Flash |
| Weights | Open weights + tech report on Hugging Face (MIT) |
DeepSeek also claims KV cache roughly 1/4 HBM and 1/8 SSD storage versus the previous generation - vendor-reported ratios, useful for capacity planning, not an independent audit.
Watchers have noted that hanging .1 on a wholly new base is unusual naming. DeepSeek itself does not explain "why not V5." Attribute that as observer framing; ship ops from the docs, not the nickname debate. For a separate frontier bake-off across other labs, see GPT-6 Astra vs Claude Fable 5.1 vs Gemini 3.8 Flash - that is a different post, not this routing checklist.
For product context, see the DeepSeek directory page.
What should API callers send after Sept 10?
| Model ID | Resolves to (as-of 2026-09-13) | Pricing class | Notes |
|---|---|---|---|
deepseek-flash | DeepSeek-V4.1-Flash | Flash rates | Canonical ID |
deepseek-v4-flash | V4.1-Flash (compat) | Flash rates | Prior V4-Flash retired; temporary |
deepseek-v4-flash-vision-exp | V4.1-Flash (compat) | Flash rates | Vision-Exp retired; temporary |
deepseek-v4-pro | DeepSeek-V4-Pro-0813 on live pricing table; news260910 still says remap to V4.1-Flash at Flash rates from 04:00 UTC Sep 14 until V4.1-Pro | Pro rates while separately served; Flash rates if/when news remap applies | CONFLICT - see next sections |
Other ops facts from the pricing table:
- Concurrency limits: Flash 2500 / Pro 500
- Vision: Flash yes; Pro "Not supported" on the published features row
- Thinking mode: non-thinking + thinking (default) for both listed models
Migrate new callers to deepseek-flash. Leave legacy Flash aliases only as temporary bridges.
What are the published Flash rates and peak hours?
New Flash pricing took effect at 04:00 UTC on September 10, 2026. As of a September 13, 2026 fetch of Models & Pricing, published Flash rates per 1M tokens look like this:
| Component | Off-peak | Peak |
|---|---|---|
| Cache-hit input | $0.003 | $0.006 |
| Cache-miss input | $0.15 | $0.30 |
| Output | $0.60 | $1.20 |
The Pro column is still listed separately as DeepSeek-V4-Pro-0813 with higher rates (for example cache-hit $0.022 / $0.044 and output $1.98 / $3.96 off-peak / peak). Pricing may change - recheck the live table before you lock unit economics.
Peak hours (UTC, Mon-Fri): 01:00-04:00 and 06:00-10:00. Off-peak is half of peak. Chinese docs express the same windows as Beijing Mon-Fri 9:00-12:00 and 14:00-18:00.
For broader tier / cost routing habits beyond DeepSeek, see how to pick the right AI model for every task. DeepSeek also publishes instruct benchmarks on its Change Log and HF card (for example GPQA Diamond, Codeforces, Terminal-Bench). Treat those as vendor-reported, not an independent bake-off against other labs.
What do official pages disagree about for Sept 14?
This is the part that breaks routers if you flatten it.
Position A - Sep 10 news post (still live as of 2026-09-13): Starting at 04:00 UTC (12:00 Beijing) on September 14, 2026, all deepseek-v4-pro requests route to V4.1 Flash at V4.1 Flash rates until V4.1-Pro launches. Same claim appears on DeepSeek's site news mirror.
Position B - live ops docs (Models & Pricing, First API Call, Change Log EN+ZH, as-of 2026-09-13): After user demand, DeepSeek will continue providing API services for DeepSeek V4 Pro after September 14, 2026, with billing method remaining unchanged. Further notice if that changes.
These are both first-party DeepSeek pages. They disagree. Prefer live pricing / changelog / First API Call footnotes for current operational policy, and keep news260910 as the historical announcement that may or may not still apply on flip morning. Secondary blogs that only quote the remap may be stale - cite DeepSeek first-party pages, then re-verify.
Re-verify flag (Sep 14 morning): screenshot pricing, First API Call, Change Log, and news260910 before and after 04:00 UTC / 12:00 Beijing. Do not hard-code a single remap outcome until those live footnotes agree.
Sept 14 routing checklist for production clients
Use this checklist only with sourced steps - no invented outcomes.
- Re-read live footnotes on Sep 13-14 morning. Compare
quick_start/pricing, First API Call, and Change Log against news260910. Screenshot what you see. - If live docs still say Pro continues: keep
deepseek-v4-proonly if you need the Pro-0813 path, Pro rate card, or the no-vision Pro column. Expect no automatic remap until DeepSeek posts a new notice. - If the news260910 remap is reinstated: after 04:00 UTC / 12:00 Beijing on Sep 14, re-test evals, latency, vision, concurrency (2500 vs 500), and billing before that clock. Remapped Pro traffic would then bill at Flash rates until V4.1-Pro.
- Migrate callers to canonical
deepseek-flash. Treatdeepseek-v4-flashanddeepseek-v4-flash-vision-expas temporary compatibility routes. - Schedule flexible jobs off-peak when you can. Do not invent savings percentages unless you compute them from the published rate cards and label the arithmetic.
When is Flash the right default vs keeping a separate Pro path?
Prefer Flash when your traffic is multimodal, agentic, cost-sensitive, or concurrency-hungry. DeepSeek documents Flash as the new-family API ID (deepseek-flash) with vision and a 2500 concurrency cap, versus Pro-0813 at 500 and no vision on the published features row - while live footnotes still list Pro as a separately served path.
Keep an explicit Pro path only while Pro-0813 remains separately served and your regression suite still requires that path (behavior, rate card, or no-vision column). Revisit the keep-or-drop decision when V4.1-Pro is announced - ship date, params, and price are not published; do not invent them.
FAQ
Are legacy deepseek-v4-flash and deepseek-v4-flash-vision-exp names safe to leave in place?
Temporarily, yes - they route to V4.1 Flash at Flash price while DeepSeek keeps the compatibility aliases. They are not a permanent contract. Migrate to deepseek-flash so a later retirement of the aliases does not surprise you.
If Pro traffic remaps (or continues), which rate card applies - Flash or Pro?
If Pro continues as a separate service (live footnote path), the Pro rate card applies. If the news remap is reinstated, remapped deepseek-v4-pro traffic would bill at Flash rates until V4.1-Pro. Your invoice follows whichever policy is live at request time - verify the footnotes, then re-test a billing sample.
Are Flash peak and off-peak windows the same in UTC and Beijing docs?
Yes - they describe the same Mon-Fri windows. English Models & Pricing lists peak as 01:00-04:00 and 06:00-10:00 UTC; Chinese docs list 9:00-12:00 and 14:00-18:00 Beijing. Off-peak is published as half of peak. Confirm the live table before you schedule batch jobs.
Is a V4.1-Pro ship date, parameter count, or price published yet?
No. DeepSeek's news post mentions V4.1-Pro only as a future launch after which any Pro-to-Flash remap would end. Live docs do not publish ship date, params, or price - do not invent them when you plan capacity or unit economics.
Do I need a separate vision model ID after V4.1 Flash?
No. Native vision is on Flash itself, and the retired deepseek-v4-flash-vision-exp alias temporarily routes to V4.1 Flash at Flash rates. New callers should use deepseek-flash for multimodal traffic; keep the vision-exp string only as a short-lived compatibility bridge.
Next step
Pin production callers to deepseek-flash now. On the morning of September 14, screenshot the live Pro footnotes on Models & Pricing, Change Log, and First API Call (and compare news260910), then decide Pro keep-or-drop from your regression needs - not from a secondary blog that only quotes the Sep 10 news remap.
Flash rates, Pro policy, and peak windows can change. Confirm current numbers and footnotes on DeepSeek's docs before you ship.
