What is DeepSeek V4 Pro?
DeepSeek V4 Pro is the flagship text model of DeepSeek’s V4 generation, shown as a preview on 24 April 2026 and made generally available across the app, the web interface and the API on 13 August 2026. Its official API identifier is deepseek-v4-pro, and the general-availability weights are published in the Hugging Face repository DeepSeek-V4-Pro-0813 under the MIT licence. The published scale is 1.6 trillion total parameters with 49 billion active, a mixture-of-experts layout that DeepSeek says uses a hybrid attention design built for long context; at a 1M-token context the provider states that V4 Pro needs 27% of the per-token inference FLOPs and 10% of the KV cache that DeepSeek-V3.2 needed for the same work. The documented context length is 1M tokens and the documented maximum output is 384K tokens. A great deal of what has been written about this model is out of date, and the reason is a genuine conflict between two official pages rather than a rumour. On 10 September 2026 DeepSeek’s announcement site said V4 Pro would be phased out, with requests routed to V4.1 Flash from 04:00 UTC on 14 September at V4.1 Flash rates until a V4.1 Pro arrived. On the same day the official API change log said the opposite: that in response to user demand DeepSeek had decided to continue providing V4 Pro after 14 September, with billing unchanged. Both sentences are official, both come from the provider, and both are still published, so this page reports them by source and date rather than choosing one. What settles the current state is not either sentence but three things you can re-check yourself on the day you read this. The official pricing table still carries a deepseek-v4-pro column of its own, labelled with the model version DeepSeek-V4-Pro-0813 and a concurrency limit of 500. The official model list returns exactly two identifiers, deepseek-flash and deepseek-v4-pro. The API reference lists those same two as the permitted values of the model parameter. A model that had actually been withdrawn would not survive all three checks at once. Two cautions belong with the figures above. DeepSeek attributes the efficiency numbers to V4 Pro specifically, so they cannot be carried over to any other member of the family. And the specification cards below are DeepSeek describing DeepSeek V4 Pro, while the controls in the chat box are what this site switches on — the two sets are not the same, and they are kept apart throughout this page.
What DeepSeek V4 Pro is worth using for
Check whether a retirement notice you read is still current
If you are deciding whether to build on this model, the useful question is not what any one article says but what the provider currently serves, and V4 Pro is the rare case where a retirement notice exists in writing and was then reversed. Two official pages published on the same day, 10 September 2026, disagree with each other. The announcement site says V4 Pro would be phased out and its requests routed to V4.1 Flash from 14 September. The API change log says DeepSeek decided to keep providing V4 Pro after that date with billing unchanged. Read them in the order they were written and neither is a mystery: one describes a plan, the other describes the decision that replaced it, and the change log is the page that governs API behaviour. What makes the question answerable rather than a matter of trust is that three of the checks are mechanical. Open the pricing table and look for a deepseek-v4-pro column carrying the model version DeepSeek-V4-Pro-0813 and a concurrency limit of 500. Call the model list and count the identifiers, of which there are exactly two. Read the model parameter in the API reference, whose permitted values are the same two. If all three still hold, the model is current, whatever a page dated earlier says. The shape of the answer is also the honest one for the two older names in the same family, deepseek-v4-flash and deepseek-v4-flash-vision-exp, which were retired on 10 September and are still accepted only as routes to V4.1 Flash — a different arrangement, covered in detail by their own pages on this site.
I read an article saying DeepSeek V4 Pro was retired and that its requests now go to another model. Walk me through how to check that claim myself: which official pages to open, what to look for on each one, and which result on each would change the conclusion. Then tell me explicitly what would have to be true for the article to be right.
Plan a long document around the documented context, not around this chat box
V4 Pro is documented at a 1M-token context with a 384K-token maximum output, and those two numbers are a specification of the API rather than a description of one conversation here. The reason to care about the context figure is not only its size but what the provider says it costs to use: at 1M tokens DeepSeek states that V4 Pro needs 27% of the per-token inference FLOPs and 10% of the KV cache that DeepSeek-V3.2 required for the same work, both of which are prefill and memory costs that decide whether a long-context job is affordable at all. That efficiency claim is attributed to V4 Pro by name and should not be quoted for the rest of the family. When you plan the work, keep three ceilings separate. DeepSeek’s API allows up to 384K output tokens, and if you leave the parameter unset the default depends on the mode: 8K without thinking, 64K with thinking, and 128K when the effort is set to its maximum. This chat box, by contrast, requests at most 4,096 output tokens per reply and publishes no context length at all, so the interface shows no context number in it. A long deliverable therefore has to be produced in parts, with the shared context carried forward between them rather than assumed.
I need to work through a long specification and produce a long written deliverable from it. Ask me how long the source is and how long each finished section needs to be. Then tell me how to split the job, what to carry forward between the parts, and where the output ceiling of this chat box will force a change of plan.
Read the benchmark table with its footnote attached
The general-availability model card publishes one table with eleven rows and eight columns, and the column that describes this model is the first one. Terminal Bench 2.1 is 87.9 against 82.7 for the V4 Flash release of 31 July 2026; DeepSWE is 62.7 against 54.4; NL2Repo is 61.5 against 54.2; Cybergym is 83.3 against 76.7. Every row on that card which carries both columns moves the same way, and that is the safest single sentence to take from it. Three things have to travel with the numbers. First, the evaluation condition is announced for one column only: the card’s footnote says the code-agent tasks for DeepSeek-V4-Pro-0813 were run with the minimal mode of DeepSeek Harness at the maximum reasoning effort, with temperature at 1.0 and top_p at 0.95, and no other column carries a matching statement, so their conditions are not announced as identical. Second, two rows are marked with a dagger, which the card defines as internal test sets rather than public benchmarks. Third, the table is not a claim that this model leads everything on the page. On Terminal Bench 2.1 it is ahead of both V4 Flash releases and still behind Kimi K3 at 88.3, with Fable-5 at 88.0 just above it, so the honest reading is a strong agentic profile rather than uniform dominance. One row is deliberately left out of the table below rather than shipped: the Agents’ Last Exam figures appear in two change-log entries, and the later entry adds that for that evaluation a text model ignored the multimodal parts of the questions, a clause the earlier entry does not repeat.
Here is an official benchmark table with several model columns and a footnote attached to only one of them. Compare the first column with the column I name, give me the size of every gap, and list the reasons the two columns might not share an evaluation condition. Then point out the rows where the first column is not the highest, because those are the ones I want to read about before trusting a summary.
Keep the official ceilings apart from this chat box
Two sets of numbers describe this model and they answer different questions. Officially the context length is 1M tokens, the maximum output is 384K tokens, thinking is on by default with three effort levels, and image understanding is explicitly not supported. In this workspace the catalog entry behind the page declares reasoning and vision as unsupported, so the composer offers no thinking switch and no image input, and no context length is published at all, which is why the interface shows no context figure. Absent here is not the same as absent in the model, and it is not the same as broken either. What this site does set is a maximum of 4,096 output tokens per reply, a platform setting well below the official ceiling, so anything longer has to be asked for in parts. Attachments accept PDF files up to 24 MB and are converted to text before the model sees them, which means a scanned page the extractor cannot read fails outright instead of producing an empty answer. There is no web search control here either. Read the specification cards as DeepSeek describing DeepSeek V4 Pro, and the controls in the chat box as what this site actually switches on.
I want to use this chat box for a task that needs long outputs and a lot of source material. Ask me what the task is and how long the output should be. Then tell me which parts of it are limited by this interface rather than by the model, how to split the work around the per-reply ceiling, and which parts I should move to a different tool entirely.
Official DeepSeek V4 Pro specifications
- Developer
- DeepSeek
- API model name
- deepseek-v4-pro (one of the two identifiers on the official model list)
- Model version
- DeepSeek-V4-Pro-0813
- Official context length
- 1M tokens
- Official maximum output
- 384K tokens
- Thinking mode
- Non-thinking and thinking, on by default; effort low, high or max
DeepSeek V4 Pro and DeepSeek V4 Flash side by side
Every value below comes from one table, the benchmark table inside DeepSeek’s DeepSeek-V4-Pro-0813 model card, checked on 23 September 2026, so a single official page carries both columns. The left column is this page’s model and the right column is the V4 Flash release of 31 July 2026, which is the model the deepseek-v4-flash identifier pointed at before it was retired; it is not DeepSeek V4.1 Flash, because that release publishes no comparable scores on that card. Rows are matched by benchmark name, never by position. The two columns do not share an announced evaluation condition: the card’s footnote states that the code-agent tasks for DeepSeek-V4-Pro-0813 were run with the minimal mode of DeepSeek Harness at the maximum reasoning effort, temperature at 1.0 and top_p at 0.95, and it attaches no such statement to any other column. Two rows are marked with a dagger, which the card defines as internal test sets rather than public benchmarks. One row that appears on the card is left out here on purpose: the Agents’ Last Exam figures are accompanied elsewhere by a note that a text model ignored the multimodal parts of those questions, so putting the two numbers in one row would compare unlike things. Read this as a comparison of two published releases from one provider, not as a promise about the quality or speed of a conversation on this site.
| Provider benchmark | DeepSeek V4 Pro | DeepSeek V4 Flash |
|---|---|---|
| HLE (wo / w tools) | 42.7 / 60.0 | 37.8 / 51.5 |
| Terminal Bench 2.1 | 87.9 | 82.7 |
| NL2Repo | 61.5 | 54.2 |
| Cybergym | 83.3 | 76.7 |
| DeepSWE | 62.7 | 54.4 |
| Toolathlon-Verified | 74.1 | 70.3 |
| AutomationBench (Public) | 31.8 | 25.1 |
| DSBench-FullStack † | 71.1 | 68.7 |
| DSBench-Hard † | 67.2 | 59.6 |
How to use DeepSeek V4 Pro online
Choose the model
This page opens with DeepSeek V4 Pro selected. The model navigation on the left and the selector in the chat box stay in sync, and switching models keeps the current conversation. Start a new chat when you want a clean context.
Write the task and the output format
State the goal, the constraints and the shape of the answer you want. This box exposes no thinking switch for this model, so there is no reasoning level to tune and a tightly scoped instruction is the main lever you have. Any prompt in this guide can be sent as it stands or edited first.
Sign in, send and check the cost
Sending a message needs an account with available points. The composer reads a reference price from the live model catalog, and a finished request reports what it actually cost. Follow up with corrections rather than starting over.
What DeepSeek V4 Pro costs
DeepSeek prices V4 Pro on a peak and off-peak schedule, and the off-peak rate is exactly half the peak rate. Per million tokens in US dollars, a cache hit on input costs $0.022 off-peak and $0.044 at peak; a cache miss on input costs $0.66 off-peak and $1.32 at peak; output costs $1.98 off-peak and $3.96 at peak. The Chinese pricing page lists the same schedule in yuan, at 0.15 and 0.30 for a cached input hit, 4.5 and 9.0 for a cache miss, and 13.5 and 27.0 for output. Those yuan figures are quoted from their own page rather than converted, because a converted number is not an official price. Peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday to Friday, excluding Chinese public holidays, which in Beijing time is 09:00 to 12:00 and 14:00 to 18:00; every other hour, weekends and Chinese public holidays included, is off-peak. This schedule took effect at 16:00 UTC on 16 August 2026, and the official page adds that product prices may vary and that DeepSeek reserves the right to adjust them, so treat the figures as a snapshot dated by the information-check date below rather than as a rate that is fixed. The concurrency limit published for this model is 500. None of that is this site’s checkout price. Here you pay in points, and as of the information-check date the catalog entry lists 120 points per million input tokens, 240 per million output tokens and 2 points per million cached tokens.
View points and plans ↗DeepSeek V4 Pro FAQ
Is DeepSeek V4 Pro still available?
Yes, on the official API and on this page. The current status is settled by three checks you can repeat. The official pricing table still carries a deepseek-v4-pro column labelled with the model version DeepSeek-V4-Pro-0813 and a concurrency limit of 500. The official model list returns exactly two identifiers, deepseek-flash and deepseek-v4-pro. The API reference permits the same two values for the model parameter. On 10 September 2026 the announcement site said V4 Pro would be phased out in favour of V4.1 Flash, and the API change log published the same day said DeepSeek had decided to keep providing it with billing unchanged. Both statements are official, both are still published, and this page reports them by source and date instead of merging them.
How does DeepSeek V4 Pro compare with DeepSeek V4.1 Flash?
DeepSeek’s own announcement says that tests by multiple parties put V4.1 Flash ahead of V4 Pro on performance, cost, speed and total runtime. It is a qualitative sentence: no testers, benchmarks or figures are named with it, so this page quotes it as the provider’s own description and does not place the two models in one benchmark table. The table here instead sets V4 Pro against the V4 Flash release of 31 July 2026, which is the last release for which the official model card publishes directly comparable scores.
Can I use DeepSeek V4 Pro for free, and do I need an API key?
Browsing this page and preparing a prompt costs nothing, and no DeepSeek API key is required, because the conversation runs on your NanoBanana Pro account. Sending a message does need an account with available points, so this is not unlimited free chat here. Whether an account carries trial points depends on the offer shown inside that account. If you would rather call DeepSeek directly, the documentation pages linked below are the right starting points.
Can DeepSeek V4 Pro read images?
No, and on this point the official and the site limits agree. DeepSeek’s pricing table marks image understanding as not supported for this model, and the catalog entry behind this page marks vision as unsupported as well, so the composer offers no image upload. Attachments accept PDF files up to 24 MB and their text is extracted before the model sees them, so a scan the extractor cannot read fails rather than returning an empty answer.
Why is there no thinking switch in this chat box?
Because the catalog entry behind this page declares reasoning as unsupported, so the composer has no control to show. The official model differs here: DeepSeek turns thinking on by default and exposes three effort levels, low, high and max, with a longer default output when thinking is active than when it is not. That is a gap between the official model and this interface rather than a property of V4 Pro, and it is stated here instead of left to be discovered.
How long can an answer be?
This site requests at most 4,096 output tokens per reply. That is a workspace setting and not a limit of the model. DeepSeek documents a 1M-token context and a 384K-token maximum output for the API, where an unset maximum defaults to 8K without thinking, 64K with thinking and 128K at the highest effort level. The catalog entry here publishes no context length, so the interface shows no context figure. Ask for long deliverables in parts and carry the context forward between them.
Are the V4 Pro weights really open?
Yes. The general-availability weights are published in the Hugging Face repository DeepSeek-V4-Pro-0813 under the MIT licence, and DeepSeek’s own collection lists that repository alongside the rest of the V4 family. The published scale is 1.6 trillion total parameters with 49 billion active. The card states that this release is built on the preview model structure with a DSpark speculative-decoding module attached, which is why the architecture figures belong to the family technical report while the release timeline belongs to the change log.
Is NanoBanana Pro the official DeepSeek website?
No. NanoBanana Pro is an independent platform offering a shared workspace for several AI models. DeepSeek develops and serves DeepSeek V4 Pro; the specifications, benchmarks and prices quoted on this page come from the official sources linked below, and DeepSeek does not sponsor or endorse this page.
Sources and model information
Information checked
Start with the task you run most often
Give it the goal, the constraints and the format you need, and keep the inputs within what this chat box can actually take.
Chat with DeepSeek V4 Pro ↑