DeepSeek V4.1 Flash is the current cheap API seat: $0.15 input / $0.60 output per million tokens off-peak. You call it as deepseek-flash. Vision is generally available on that same ID — the old vision-exp preview is gone. Peak hours double the rate. V4 Flash is retired. V4 Pro is still on the price list, at $0.66/$1.98 off-peak, and it still cannot see images.
If you already send DeepSeek traffic, change the model string and re-price the job. The rest of this page is the rate card, a worked monthly bill, and who should stay on Pro. Prices below are from DeepSeek's own docs, checked 20 September 2026.
Key takeaways
- Call
deepseek-flash. Legacy IDsdeepseek-v4-flashanddeepseek-v4-flash-vision-expstill route here and bill at Flash. Do not keep the old names as a plan. - Off-peak list is $0.15 / $0.60 per 1M tokens (cache-miss input / output). Cache-hit input is $0.003. Peak is exactly 2×. Peak windows: 01:00–04:00 and 06:00–10:00 UTC, Monday–Friday, excluding Chinese public holidays. Weekends are all off-peak.
- Vision is GA on Flash. Pro still has no vision. 1M context, 384K max output, thinking on by default, 2,500 concurrent requests vs Pro's 500.
- V4 Pro is still listed at $0.66/$1.98 off-peak after 14 September. DeepSeek first said it would force-route Pro onto Flash; the live pricing table did not. Check the ID you actually send.
What shipped on 10 September
DeepSeek cut over to V4.1 Flash at 04:00 UTC on 10 September 2026. The model is a 552-billion-parameter mixture-of-experts with an encoder–decoder split: about 8 billion active parameters on input, 16 billion on output. DeepSeek says the KV cache — the memory the model keeps of what you already sent, which is what a cache-hit bill is paying for — now needs a quarter of the HBM and an eighth of the SSD of the previous Flash. That is why cache-hit input dropped to $0.003 per million tokens off-peak.
The company also says third-party tests put Flash ahead of V4 Pro on quality, cost, speed and total runtime. Treat that as DeepSeek's claim, not a Pickurai bench. What you can verify without a bench: Flash is cheaper, sees images, and has five times the concurrency cap.
Weights are on Hugging Face as deepseek-ai/DeepSeek-V4.1-Flash if you want to self-host. The hosted API still sends prompts to DeepSeek's servers in China. That split has not changed. For the jurisdiction answer, use are Chinese AI models safe — the weights are software; the hosted API is the risk.
The IDs, and the Pro story
Three strings now land on V4.1 Flash:
deepseek-flash— the current name. Use this.deepseek-v4-flash— retired model, temporarily routed and billed as Flash.deepseek-v4-flash-vision-exp— same. The preview is dead; vision is just Flash now.
Base URL is still https://api.deepseek.com in OpenAI format, or https://api.deepseek.com/anthropic if you speak Anthropic. Thinking mode is on by default; you can turn it off. JSON, tool calls, and fill-in-the-middle (non-thinking only) are all listed as supported.
V4 Pro is the messy row. The 10 September release note said that from 04:00 UTC on 14 September, every deepseek-v4-pro request would be served by Flash and billed at Flash rates until V4.1 Pro launches. Then DeepSeek wrote that it would keep V4 Pro on the API after that date, same billing, after user demand. As of 20 September the official pricing table still lists both models at different prices. Do not assume a silent remap. Log the model in the response and look at the invoice.
The rate card
All figures are USD per million tokens. Off-peak is half of peak. Cache-hit is a separate input line — output never gets that discount.
| Line | Flash off-peak | Flash peak | Pro off-peak | Pro 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 |
Source: DeepSeek Models & Pricing, checked 20 September 2026. Previous Flash (Aug 21 list) was $0.22/$0.66 off-peak on a miss. Flash concurrency 2,500; Pro 500. Vision: Flash yes, Pro no.
Against the August Flash list, a miss is cheaper on input (32% down) and a little cheaper on output (9% down). The cache-hit line is the real cut: $0.007 → $0.003 off-peak. Agent loops that resend the same system prompt will feel that first.
A real monthly bill
Same workload we use elsewhere on the site: 25,000 requests a month, about 1,500 input tokens and 800 output tokens each. That is 37.5 million input tokens and 20 million output tokens. Cache-miss, unless the row says otherwise. Dated September 2026.
| Model | List (in/out) | This workload | Best for |
|---|---|---|---|
| V4.1 Flash, off-peak miss | $0.15 / $0.60 | ≈ $18 | Default DeepSeek API seat. |
| V4.1 Flash, off-peak, input cached | $0.003 / $0.60 | ≈ $12 | Agents that resend a long prefix. |
| V4.1 Flash, peak miss | $0.30 / $1.20 | ≈ $35 | Still cheaper than Pro off-peak. |
| V4 Pro, off-peak | $0.66 / $1.98 | ≈ $64 | Only if you measured Pro better on your prompts. |
| GPT-5.6 Luna | $0.20 / $1.20 | ≈ $32 | Cheapest current OpenAI chat tier. |
| Gemini 3.8 Flash | $0.75 / $3.75 | ≈ $103 | Google stack, intro rate through 31 Dec 2026. |
| Claude Sonnet 5 | $2 / $10 | ≈ $275 | Western mid-flagship. Files and Cowork live here, not on Flash. |
Flash off-peak miss: 37.5 × $0.15 + 20 × $0.60 = $17.63. Cached input: 37.5 × $0.003 + 20 × $0.60 = $12.11. Run your own mix in the token calculator — it already bills this alias as V4.1 Flash at the off-peak miss rate.
Two reads fall out of that table. First: Flash off-peak is still the cheapest capable API we track for this job, including Luna. Second: if your traffic sits in the UTC morning peak, Flash still undercuts Pro off-peak. The peak penalty is real (~$18 → ~$35) but it does not throw you onto a Western invoice.
This is the same shape as the 2026 API price war: Chinese labs win on volume. Flash just moved the floor again. The older roundup in Chinese AI models pricing still has August Flash numbers ($0.14/$0.28 on an even older list). Use this page for DeepSeek. Use that page for the rest of the Chinese ladder.
Flash or Pro — who should pick which
Pick Flash if you are starting a DeepSeek integration, if you need images in the same call, if you run agents that hit the cache, or if the bill is the point. That is most API users. New code should send deepseek-flash. The consumer chat app is still free; DeepSeek does not sell a Plus plan. The API is how they charge. Setup is in the DeepSeek API beginner's guide — swap the model name to deepseek-flash.
Stay on Pro only if you already A/B tested it on your prompts and Flash lost, and if the ID still returns Pro on your key. Do not stay because the name says Pro. DeepSeek is telling you Flash is the better model. Prove it on ten tasks before you pay 4× on input and 3× on output.
Do not pick hosted DeepSeek for confidential or regulated work. Same answer as in DeepSeek vs ChatGPT and Kimi vs DeepSeek: China hosting, training-on-your-data by default unless you opt out, tighter political filters. Self-host the open weights, or use Claude, which does not train on paid conversations by default.
Flash is the default DeepSeek API now. $0.15/$0.60 off-peak, vision included, old IDs already point here. Pro is the exception you have to justify.
If you are choosing a model this week and you are not already locked to DeepSeek, start at the wizard. Six questions. It will not send a regulated workload to a China-hosted API unless that is what you asked for.