by DeepSeek
DeepSeek's flagship open-weight reasoning model: a 1.6T-parameter MoE with 1M context, an MIT license and weekday peak and off-peak API rates.
Pick V4-Pro when you want coding and agent performance from weights you can download and run privately, or hosted rates that halve outside weekday peak hours. Skip it if you need image or audio input. Artificial Analysis, scoring it 36 on its Intelligence Index, still calls its per-token price expensive for an open-weight model of this size.
| Job | Input | Output | Total |
|---|---|---|---|
| Summarise a 20-page PDF | $0.040 | $0.0040 | $0.044 |
| Support reply | $0.0026 | $0.0012 | $0.0038 |
| One coding agent run | $0.264 | $0.079 | $0.343 |
Budgets: 20-page PDF = 30k in / 1k out · Support reply = 2k in / 300 out · Coding agent run = 200k in / 20k out. Computed from the vendor's per-token prices at render time; cached-input discounts are not applied.
Prices are per 1M tokens and depend on the time of day. Peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday to Friday, excluding Chinese public holidays: $1.32 cache-miss input, $0.044 cache-hit input and $3.96 output. Every other hour, including weekends and Chinese public holidays, is off-peak at half those rates: $0.66, $0.022 and $1.98. This schedule took effect at 16:00 UTC on August 16, 2026, replacing the earlier flat $1.74 and $3.48 list price and the 75% launch promotion that ended May 31, 2026. DeepSeek says prices may change. Artificial Analysis lists the peak rate. For comparison, the records for Claude Opus 4.8 list $5 and $25, GPT-5.5 $5 and $30, and Qwen3.7-Max $2.50 and $7.50 per 1M input and output tokens. Self-hosting the MIT weights carries infrastructure cost only.
DeepSeek-V4-Pro is DeepSeek's flagship open-weight Mixture-of-Experts model: 1.6 trillion parameters, 49 billion active per token, a 1M-token context window and an MIT license. It reached general availability on August 13, 2026, after an April preview. It takes text input only and offers low, high and max reasoning effort through the API.
Reads
Writes
Reasoning Effort Control
Thinking and non-thinking modes, with low, high and max effort levels added at the general release.
1 Million Token Context Window
Fits a full codebase or document set in one request, with the output cap set at 384K tokens.
MIT License Open Weights
Full weights on Hugging Face at deepseek-ai/DeepSeek-V4-Pro with no restrictions on commercial use or fine-tuning.
OpenAI and Anthropic API Formats
One API serves OpenAI-format requests, the Responses API and an Anthropic-format endpoint, with JSON output and tool calls.
Peak and Off-Peak Pricing
Weekday peak windows of seven hours UTC carry double the rate of every other hour, weekends and Chinese public holidays.
Priced around the middle of the 106 GA models with a published price (rank 55), in the top third on SWE-bench Verified (rank 12 of 39), and one of 90 whose vendor states it does not train on customer data.
| Measure | This model | Rank | Median |
|---|---|---|---|
| Blended price (3:1)$ per 1M tokens | $1.98/M | #55 / 106 | peer median $1.73/M |
| Output speedtokens/s | 92 tok/s | #31 / 59 | peer median 94 tok/s |
| SWE-bench Verified% solved | 80.6% | #12 / 39 | peer median 76.5% |
| GPQA Diamond% correct | 90.1% | #22 / 63 | peer median 87.8% |
Ranks are against GA models on Hokai that publish the same figure; ties share a rank.
A benchmark is an exam, not the job. Scores transfer unevenly between tasks, so weigh the one closest to your workload and read every figure with its source.
| Model | Lab | Context | In / out per 1M | Reads and writes |
|---|---|---|---|---|
| DeepSeek | 1M | $1.32 / $3.96 | text → text | |
| Alibaba Cloud | 1M | $2.50 / $7.50 | text → text | |
| Anthropic | 1M | $5 / $25 | text, image, pdf → text | |
| Tencent | 1M | $0.07 / $0.281 | text, image → text |
Training data
Pre-trained on 33 trillion tokens with emphasis on long documents, code, math, scientific text, and agentic execution traces. Post-training uses two stages: SFT plus GRPO domain cultivation, then unified model consolidation via on-policy distillation. Sensitive personal data screened per April 2026 model card. Trained on Huawei Ascend 950PR chips.
| Name | Endpoint | Auth |
|---|---|---|
| DeepSeek API | https://api.deepseek.com/v1/chat/completions | API Key |
| AWS Bedrock | https://aws.amazon.com/bedrock/ | AWS IAM |
| Google Vertex AI | https://cloud.google.com/vertex-ai | GCP IAM |
| Azure AI Foundry | https://azure.microsoft.com/en-us/products/ai-foundry | Azure AD |
| Fireworks AI | https://fireworks.ai/models/deepseek-ai/deepseek-v4-pro | API Key |
| Together.ai | https://www.together.ai/ | API Key |
| DeepInfra | https://deepinfra.com/ | API Key |
| OpenRouter | https://openrouter.ai/deepseek/deepseek-v4-pro | API Key |
Set reasoning mode explicitly per request
Include reasoning_effort: 'high' or 'max' in the API call for complex tasks; omit for simple lookups to save tokens.
Think Max mode costs more context budget tokens; Non-think is 3-5x faster for routine tasks.
Use OpenAI-compatible SDK with base_url swap
client = OpenAI(base_url='https://api.deepseek.com/v1', api_key=DS_KEY); model='deepseek-v4-pro'
DeepSeek API is fully OpenAI ChatCompletions-compatible; no SDK changes needed beyond base_url and model name.
Think Max requires 384K token context budget
Workaround: Set context window to at least 384K tokens when using Think Max mode, otherwise the model falls back to shorter reasoning chains.
Rates double during weekday peak windows
Workaround: Schedule non-urgent batch jobs outside 01:00-04:00 and 06:00-10:00 UTC on weekdays, or on weekends and Chinese public holidays, to pay the off-peak rate.
DeepSeek-V4-Pro-0813 · 13 Aug 2026
General availability. Adds low, high and max reasoning effort and native OpenAI Responses API support. Peak and off-peak API pricing took effect 2026-08-16 16:00 UTC.
deepseek-v4-pro · 24 Apr 2026
Initial preview release. SWE-bench 80.6% (Think Max), LiveCodeBench 93.5, Codeforces 3206. MIT license, 1M context, 384K max output. Three reasoning modes.
deepseek-v3.2 · 01 Jan 2026
Prior generation. SWE-bench 67.8%, LiveCodeBench 74.1, 128K context. Predecessor to V4 series.
During the weekday peak windows it costs $1.32 per 1M cache-miss input tokens, $0.044 for cache-hit input and $3.96 for output. All other hours are off-peak at half that: $0.66, $0.022 and $1.98. The schedule started on August 16, 2026. Self-hosting the open weights has no per-token fee beyond GPU costs.
New API accounts receive 5 million free tokens to start, with no separate free tier beyond that credit. Teams that want ongoing free access can instead self-host the MIT-licensed open weights, though that requires a multi-GPU cluster rather than a hosted plan.
Qwen3.7-Max is a hosted rival with a 1M context window and a higher listed GPQA Diamond score, but proprietary weights. Claude Opus 4.8 and the GPT flagship add image input and list higher SWE-bench Verified scores at higher per-token prices. DeepSeek-V4-Flash is the same family with image input.
Claude Opus 4.8 lists 88.6% on SWE-bench Verified and 93.6% on GPQA Diamond against V4-Pro's preview-card 80.6% and 90.1%, and it accepts images and PDFs where V4-Pro takes text only. V4-Pro counters with MIT-licensed downloadable weights and API rates well below the Opus 4.8 record's listed prices.
Create an API key on the DeepSeek platform and point an OpenAI-compatible SDK at https://api.deepseek.com with the model name deepseek-v4-pro. An Anthropic-format endpoint is available at https://api.deepseek.com/anthropic. Choose thinking or non-thinking mode, and set reasoning effort to low, high or max. Self-hosters pull the weights from Hugging Face instead.
DeepSeek-V4-Pro is the flagship model from DeepSeek, the Chinese AI lab founded in 2023 by High-Flyer hedge fund CEO Liang Wenfeng. It first appeared as a preview on April 24, 2026 under the MIT license, and DeepSeek announced general availability on August 13, 2026 with a new reasoning-effort control and native OpenAI Responses API support. The current API build is labelled DeepSeek-V4-Pro-0813. It is a Mixture-of-Experts transformer with 1.6 trillion total parameters and 49 billion active per token. According to the April model card it was pre-trained on 33 trillion tokens and trained on Huawei Ascend 950PR chips rather than Nvidia hardware. Its smaller sibling, DeepSeek-V4-Flash, trades capability for throughput and, unlike Pro, accepts images.
The context window is 1,048,576 tokens with up to 384,000 tokens of output per request, per DeepSeek's API docs. The API accepts OpenAI-format and Anthropic-format requests, and supports JSON output, tool calls, the Responses API, chat prefix completion (beta) and FIM completion in non-thinking mode (beta). Thinking and non-thinking modes are both available, and DeepSeek recommends low reasoning effort for simple tasks, high for daily agent workflows and max for complex ones. Pro does not support vision: input is text only, so image, audio and video work belongs on a sibling or a rival such as Gemini 3.1 Pro. Coding agents such as OpenCode and Claude Code are among the agent integrations DeepSeek documents for its API.
The benchmark scores stored on this page come from DeepSeek's April 27, 2026 model card for the preview checkpoint: 80.6% on SWE-bench Verified in Think Max mode, 90.1% on GPQA Diamond, 37.7% on Humanity's Last Exam, 93.5 on LiveCodeBench and a Codeforces rating of 3,206. DeepSeek published a new benchmark table for the 0813 build as an image on its announcement page, and those figures have not been transcribed here, so treat the stored scores as the preview's. Artificial Analysis, an independent tester, scores the 0813 Max configuration at 36 on its Intelligence Index (v4.3.2) with output speed of about 92 tokens per second, and says the model generated 160 million tokens running the index, which it calls somewhat verbose.
Teams running agentic coding pipelines, long-document analysis or private inference have the clearest use for it. The MIT license allows fine-tuning, air-gapped deployment and redistribution, which no closed rival such as Claude Opus 4.8 or GPT-5.5 offers. The other side of the ledger is real. Both of those rivals accept images and list higher SWE-bench Verified scores on their records, and Qwen3.7-Max lists a higher GPQA Diamond score. Artificial Analysis also rates V4-Pro's per-token price as expensive next to other open-weight models of similar size, so the case for it rests on the open weights and the off-peak rates rather than on being the cheapest model available. The concurrency limit on DeepSeek's own API is 500 for Pro, and self-hosting needs roughly 865GB of FP4 and FP8 weights and a multi-GPU cluster, with 4x H200 141GB the practical minimum cited in the April card's deployment notes.
DeepSeek's published safety documentation is thinner than that of Western frontier labs. The April model card says sensitive personal data and identification information were excluded from training data, and it discloses no external red-team partners. The MIT license also means anyone can download the weights and strip the safety tuning, a different risk profile from an API-only model. The version history on this page lists the preview, the 0813 general-availability build and its predecessor DeepSeek-V3.2, whose 128K context and 67.8% SWE-bench Verified score the V4 series replaced.
Specifications on this page (context window, token prices, benchmark scores, licence) are read from DeepSeek's own model card and documentation. The verdict is Hokai's reading of them, not the provider's claim.