Checked Oct 17, 2026
DeepSeek
DeepSeek V4 Pro
Best for: Low-cost agentic coding and tool use, with MIT-licensed open weights.
- LMArena text #56 of 416
- Reasoning
- Open weights
- Reads text
- Input
- $1.32
- per 1M tokens
- Output
- $3.96
- per 1M tokens
- Cached input
- $0.044
- per 1M tokens
- Context window
- 1M
- tokens in one request
- Max output
- 375K
- tokens per reply
- Released
- Aug 13, 2026
- Weights
- Open
- MIT
Peak-hour price shown (cache miss input). Off-peak price is half: $0.66 input / $1.98 output / $0.022 cache hit. Peak = 01:00-04:00 and 06:00-10:00 UTC, Mon-Fri (excluding Chinese public holidays). Supports thinking and non-thinking modes; no vision. Context listed as 1M.
What would DeepSeek V4 Pro cost you?
Change the numbers; the chart compares all 10 featured models at the same usage.
What it's good at
From DeepSeek's own announcements and docs, linked.
- The August general release is a big agent upgrade over the April preview: DeepSeek reports 87.9 on Terminal-Bench 2.1 and 62.7 on DeepSWE. source
- The weights are on Hugging Face under the MIT license, so you can self-host it or fine-tune it for commercial use. source
- A 1M-token context, replies of up to 384K tokens, thinking and non-thinking modes, JSON output and tool calls. source
Watch out for
Limits and costs worth knowing before you build on it.
- The price shown is the peak rate. Off-peak, outside 01:00–04:00 and 06:00–10:00 UTC on weekdays, it's half price.
- Text only: no image input.
- DeepSeek doesn't state a knowledge cutoff.
- Running it yourself takes a multi-GPU server: the weights are about 893 GB.
Scores
Benchmark numbers are as published by DeepSeek; we haven't rerun them. Makers use different setups, so don't compare one maker's numbers with another's. The leaderboard rank comes from public blind votes.
| Measure | Result | Source |
|---|---|---|
| LMArena text rank | #56 (1465) | Oct 15, 2026 |
| Terminal-Bench 2.1 (agentic coding) | 87.9 | DeepSeek |
| DeepSWE (software engineering) | 62.7 | DeepSeek |
| HLE (expert reasoning, without / with tools) | 42.7 / 60.0 | DeepSeek |
How to start
Try it in the app
The quickest way to see if it suits your work. Open DeepSeek's app.
Get an API key
Sign up on DeepSeek's developer platform and add billing; you pay per token as shown above. Read the docs.
Call the model
Use this model ID in your DeepSeek API requests:
deepseek-v4-proOr run it yourself
Download the weights from Hugging Face and serve them on your own or a rented GPU (see below).
Run DeepSeek V4 Pro on your own GPU
Open weights mean you can rent a GPU by the hour and serve the model yourself: a flat hourly cost however many tokens you generate. DeepSeek gives no memory figure. The weights are about 893 GB (roughly 1.6T parameters in FP8), and DeepSeek's own example serves them on a node of four GB300 GPUs, so plan on a multi-GPU server.
| Provider | From | Starting plan | Free tier | Our rating | Link |
|---|---|---|---|---|---|
| RunPodChecked | $0.34/hr | RTX 4090 24GB (Community Cloud) | No free credits; small referral bonus after a $10 top-up | Not rated | VisitOfficial site |
| TensorDockChecked (partly unverified) | $0.35/hr | RTX 4090 24GB | No ($5 minimum deposit) | Not rated | VisitOfficial site |
| Vast.aiChecked | $0.49/hr | RTX 4090 24GB | No free credits | Not rated | VisitOfficial site |
| DigitalOcean GPU DropletsChecked | $0.76/hr | RTX 4000 Ada 20GB8 vCPU · 32 GB RAM | New-account credit; GPU use not confirmed | Not rated | VisitOfficial site |
| ModalChecked | $1.10/hr | A10 | $30/month of free credits on the Starter plan | Not rated | VisitOfficial site |
| HyperstackChecked | $1.35/hr | A100 80GB (PCIe)A100 PCIe · 28 vCPU · 120 GB RAM | No | Not rated | VisitOfficial site |
| Vultr Cloud GPUChecked | $1.67/hr | L40S 48GB (1 GPU)16 vCPU · 180 GB RAM | New-customer promo credit; GPU use not confirmed | Not rated | VisitOfficial site |
| LambdaChecked | $2.29/hr | GH200 96GB (1x instance)64 vCPU · 432 GB RAM | No | Not rated | VisitOfficial site |
Compare with the other 9
| Model | Input / 1M | Output / 1M | Context | Arena |
|---|---|---|---|---|
| gpt-oss-120bOpenAI | $0.15 | $0.60 | 128K | #215 |
| Gemini 3.8 FlashGoogle | $0.75 | $3.75 | 1M | #9 |
| DeepSeek V4 ProDeepSeek | $1.32 | $3.96 | 1M | #56 |
| Grok 4.7xAI | $2 | $6 | 500K | #89 |
| Gemini 4 ArgonGoogle | $2 | $10 | 1M | #1 |
| GPT-6.1 SolOpenAI | $2 | $10 | 1.05M | #22 |
| Claude Sonnet 5.5Anthropic | $2 | $10 | 1M | #36 |
| Claude Opus 5.5Anthropic | $4 | $20 | 1M | #3 |
| Claude Fable 5.1Anthropic | $10 | $50 | 1M | #6 |
| GPT-6 AstraOpenAI | $10 | $50 | 1.05M | #30 |
Cheapest first, using 3 input tokens per output token.
Questions
How much does DeepSeek V4 Pro cost?
DeepSeek V4 Pro costs $1.32 per million input tokens and $3.96 per million output tokens at standard API rates, and $0.044 per million cached input tokens (US dollars, checked Oct 17, 2026). It is the 3rd cheapest of the 10 models we feature.
How big is DeepSeek V4 Pro's context window?
1,000,000 tokens for the prompt, documents and conversation in one request. A single reply can be up to 384,000 tokens. How many words that holds depends on the model's tokenizer.
Is DeepSeek V4 Pro open source?
Its weights are published under the MIT license, so you can download it and run it on your own GPU or a rented one, as well as use it through hosted APIs.
When was DeepSeek V4 Pro released?
DeepSeek released it in Aug 13, 2026.
Sources
- DeepSeek pricing, checked Oct 17, 2026
- api-docs.deepseek.com/quick_start/pricing
- api-docs.deepseek.com/news/news260813
- api-docs.deepseek.com/updates
- huggingface.co/deepseek-ai/DeepSeek-V4-Pro-0813
Facts checked Oct 17, 2026. Model names belong to their makers; AskGPU isn't affiliated with DeepSeek. Spotted something out of date? Tell us at hello@askgpu.com.