DeepSeek’s API lists two models as of 9 October 2026. Call deepseek-flash for DeepSeek-V4.1-Flash, released 10 September 2026. That is the newest model, and it is the one with vision. Call deepseek-v4-pro for DeepSeek-V4-Pro-0813, which is still on the API. The old ids deepseek-v4-flash and deepseek-v4-flash-vision-exp are retired. DeepSeek still accepts them and serves V4.1-Flash at the Flash price. R1 is the January 2025 reasoning model. It is not a current API id. The prices below were read from DeepSeek’s pricing page on 9 October 2026. They use peak and off-peak rates.
What a 1,024 token cap did to thinking mode
On 9 October 2026 a client sent one editing prompt to deepseek-flash twice, through DeepSeek’s API, with a 1,024 token output cap. The sample is two requests. Thinking off returned the three-line answer in 89 output tokens. Thinking on, which DeepSeek’s pricing page calls the default, spent all 1,024 tokens on reasoning and returned no answer. Input on that call was 197 tokens. The call was metered at about $0.000644 and finished in 4.65 seconds.
Same prompt, same 1,024 token cap: thinking off vs on
Pick a setting to see what the editor received. Bars show output tokens against the 1,024 token cap.
Two requests to deepseek-flash on 9 October 2026. Input was 197 tokens on the thinking-on call.
| Setting | What the editor received | Tokens |
|---|---|---|
| Thinking off | Three lines: which id to switch to, the 10 September release date, and a refusal to say V4 Pro was retired | 89 output |
| Thinking on, default effort | No answer. The cap filled with reasoning | 197 input, 1,024 output, all of it thinking |
For a short structured edit, turn thinking off or raise the cap before you send. A 1,024 token limit left no room for the default thinking mode to finish and then write the three lines.
The thinking-off answer, in full:
Decision: Buyers still using V4 Flash or V4 Flash Vision Exp can switch to deepseek-flash, since those retired names now route to V4.1-Flash. Keep: DeepSeek released DeepSeek-V4.1-Flash on 10 September 2026. Cut: Do not publish that V4 Pro was retired, because the source only says its API service continued after 14 September 2026.
The prompt told the model to use only a short source, to add no prices or dates that were not in it, and to return exactly three lines: a buyer decision, one fact with its date, and one thing not to publish. Both calls used a fresh conversation, OpenAI chat format, and the same 1,024 token cap. deepseek-v4-pro was not in this test.
Current API models
Call deepseek-flash or deepseek-v4-pro. Those are the two current ids. A retired name still routes to V4.1-Flash. Thinking and effort are settings on the call, not extra models.
DeepSeek API ids on 9 October 2026
Search, filter by status, or click a column heading to sort.
deepseek-flash | DeepSeek-V4.1-Flash | Current | Supported | 1,000,000 tokens | 384,000 tokens | 2,500 |
deepseek-v4-pro | DeepSeek-V4-Pro-0813 | Current | Not supported | 1,000,000 tokens | 384,000 tokens | 500 |
deepseek-v4-flash | Served by V4.1-Flash, billed at the Flash price | Retired alias | As deepseek-flash | As deepseek-flash | As deepseek-flash | n/a |
deepseek-v4-flash-vision-exp | Served by V4.1-Flash, billed at the Flash price | Retired alias | As deepseek-flash | As deepseek-flash | As deepseek-flash | n/a |
| R1 (January 2025 reasoning model) | Not a current API id | Not on the pricing page | n/a | n/a | n/a | n/a |
No rows match. Clear the search or press Reset.
Source: DeepSeek models and pricing, read 9 October 2026. Both current ids support thinking and non-thinking, and thinking is the default. The 13 August 2026 changelog gives V4-Pro effort levels of low, high, and max. The June labels Non-Thinking, Thinking High, and Thinking Max are out of date.
The 10 September 2026 announcement describes V4.1-Flash as a 552-billion-parameter mixture-of-experts model that activates 8 billion parameters on input and 16 billion on output.
Prices on 9 October 2026
Rates are US dollars per 1 million tokens. Off-peak is half of peak. Peak hours on the pricing page are 01:00 to 04:00 UTC and 06:00 to 10:00 UTC, Monday through Friday, excluding Chinese public holidays. Other hours, including weekends and those holidays, are off-peak.
Price per 1 million tokens, Flash vs Pro
Choose the rate type and the time of day. The bars use one scale, so you can see the gap.
Source: DeepSeek models and pricing, read 9 October 2026.
| 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 |
The pricing page lists a concurrency limit of 2,500 for deepseek-flash and 500 for deepseek-v4-pro. Read the page again before you buy. DeepSeek says prices can change.
Cost calculator
Estimate a job at the 9 October 2026 rates. Enter tokens, not words.
Thinking tokens count as output tokens. The test call (197 input, 1,024 output, Flash, off-peak, no cache hit) comes to about $0.000644. This tool checks UTC hours and weekdays only. It does not know Chinese public holidays, which are off-peak.
Which id to call
Pick the id from the job. Then set thinking on or off for that request.
Pick the id for the job
Type a word from your job, such as image, R1, or regulated, or use a filter.
| Job | Call this id | Why |
|---|---|---|
| High-volume chat, classification, extraction, or a request that includes an image | deepseek-flash | Current Flash model. Vision is on this id. Pro does not list vision. |
| Long-document work or a hard reasoning step where you still want the Pro checkpoint | deepseek-v4-pro | Still listed, version DeepSeek-V4-Pro-0813. No vision. |
Code that already sends deepseek-v4-flash or deepseek-v4-flash-vision-exp | Change the id to deepseek-flash | Those names are retired aliases. They currently route to V4.1-Flash and bill at the Flash price. |
| A paper, a fine-tune, or an existing deployment built on R1 | Keep the R1 checkpoint you already run | R1’s public reasoning release is 20 January 2025. It is not one of the two ids on the pricing page. |
| A prompt that contains regulated data | Do not send it to a hosted API until you have confirmed where that provider stores it | A downloaded checkpoint and the hosted API are different offers. Read the model card for the checkpoint you intend to run. |
No rows match. Clear the search or press Reset.
What changed after the June guide
The June version of this page called V4-Pro the latest model and listed V4-Flash, id deepseek-v4-flash, as the current cheap sibling. Both of those statements are behind the changelog.
DeepSeek API changelog, 2026
Filter by model line, search, or click Date to reverse the order.
| What DeepSeek published | |
|---|---|
| 24 April 2026 | V4-Pro and V4-Flash on the API as deepseek-v4-pro and deepseek-v4-flash |
| 31 July 2026 | V4-Flash official release in public beta |
| 13 August 2026 | V4-Pro general availability. Version string DeepSeek-V4-Pro-0813 |
| 21 August 2026 | Experimental V4-Flash-Vision-Exp, id deepseek-v4-flash-vision-exp |
| 10 September 2026 | V4.1-Flash. Call it as deepseek-flash. V4-Flash and V4-Flash-Vision-Exp retired, with those ids routed to V4.1-Flash |
No rows match. Clear the search or press Reset.
Source: DeepSeek API changelog, read 9 October 2026.
The 10 September news post said V4-Pro requests would move to V4.1-Flash on 14 September. The changelog and the pricing page, read the same day as this draft, say V4-Pro API service continued after that date, with billing unchanged. Follow the pricing page if those two disagree.
R1 stays in the record as the model that made DeepSeek widely known. DeepSeek’s technical report estimated $5.6 million of training compute for the V3 model underneath it, and the reinforcement-learning stage was later disclosed at $294,000. No R2 entry appears on the changelog read on 9 October 2026.
A June 2026 snapshot of this guide treated $0.87 and $0.28 output rates as the permanent list prices. Those rates are not on the pricing page read on 9 October 2026. The changelog records price changes on 16 August 2026 and again on 10 September 2026.
How a team should route it
Use one provider when one model covers the work. Add a second provider when a job needs a different tradeoff, such as vision on Flash, or a checkpoint you are not willing to send to DeepSeek’s API. Each extra provider adds another key, another bill, and another place where a prompt lives in one person’s account.
TeamAI is a shared workspace for that split: prompts, agents, and access controls in one place, with one signup at app.teamai.com/signup. WebFX runs TeamAI with 500+ employees. Check the model picker before a workflow depends on a specific DeepSeek id.
As of 9 October 2026, the TeamAI model list shows DeepSeek V3 and DeepSeek R1.
Related: How to choose the right LLM, Why use more than one large language model, and What to look for in a unified AI workspace.
Frequently Asked Questions
What is the latest DeepSeek model?
DeepSeek-V4.1-Flash, released 10 September 2026. On the API you call it as deepseek-flash. DeepSeek-V4-Pro-0813 is still listed as deepseek-v4-pro. It is not the newest release.
How much does DeepSeek cost?
On 9 October 2026, Flash output was $0.60 per million tokens off-peak and $1.20 at peak. Pro output was $1.98 off-peak and $3.96 at peak. Input has separate cache-hit and cache-miss rates in the table above. Confirm the live numbers on DeepSeek's pricing page before you commit spend.
What is the difference between V4.1-Flash and V4-Pro?
V4.1-Flash is the 10 September model, id deepseek-flash, and the pricing page marks vision as supported. V4-Pro on that page is the 13 August checkpoint, id deepseek-v4-pro, and vision is not supported. Both list a 1 million token context and a 384,000 token output cap. Flash is the lower price.
Which DeepSeek model should I use?
Call deepseek-flash for volume, vision, and the current default. Call deepseek-v4-pro when you still want that Pro checkpoint. Leave R1 on deployments that already depend on it. If your code still says deepseek-v4-flash, change it to deepseek-flash so the id matches the model you are actually getting.
Is the old V4-Flash id still valid?
DeepSeek still accepts deepseek-v4-flash and deepseek-v4-flash-vision-exp. The pricing page says those models are retired and that the requests are served by V4.1-Flash at the Flash price. The id in your code is no longer the model name.
Should I send company data to the DeepSeek API?
Confirm where the provider stores prompts before the data is regulated, confidential, or covered by a customer contract. A downloaded checkpoint and the hosted API are different decisions. Read the model card for the checkpoint you intend to run.