DeepSeek V4 Pro 101: The 1.6T-Parameter Open Model at $0.435/M Input

Priya Nair

Priya Nair

AI Infrastructure Analyst

Published: August 13, 2026
DeepSeek V4 Pro Release Coverage editorial cover

TLDRA signal-based analysis of DeepSeek V4 Pro and the latest release discussion.

DeepSeek V4 Pro is an open-weights Mixture-of-Experts (MoE) language model developed by DeepSeek, with 1.6 trillion total parameters and 49 billion active per token, a 1M-token context window, and MIT-licensed weights. It is priced at $0.435 per 1M input tokens and $0.87 per 1M output tokens on the official API. First released in preview on April 24, 2026, it has since rolled out as a general-availability build tagged deepseek-v4-pro-0813, which went live across DeepSeek's API and chat on August 12, 2026.

Key Takeaways

  • DeepSeek V4 Pro is a 1.6T-parameter MoE model with 49B active parameters and a 1M-token context window, released under the MIT license.
  • Official API pricing is $0.435 per 1M input tokens (cache miss), $0.003625 on a cache hit, and $0.87 per 1M output tokens.
  • The deepseek-v4-pro-0813 GA build shipped on August 12, 2026, with DeepSeek-reported agent-benchmark gains over the preview.
  • On Artificial Analysis, the reasoning variant scored 44 on the Intelligence Index (#3 of 94) with the lowest cost-per-task in its comparison set.
  • A May 2026 NIST/CAISI evaluation placed the preview roughly 8 months behind the U.S. frontier despite strong self-reported scores.
  • The model is text-only (no vision), and its release status briefly wobbled on August 13, 2026, when some official notices were reportedly withdrawn.

What Is DeepSeek V4 Pro?

DeepSeek V4 Pro is the larger of the two models in DeepSeek's V4 series, sitting above DeepSeek V4 Flash (284B total / 13B active). Both are open-weights MoE language models built for what DeepSeek calls "cost-effective 1M context length." Pro carries 1.6 trillion total parameters with 49 billion activated per token, which makes it the largest open-weights model shipped to date by parameter count.

The model targets reasoning-heavy and coding-heavy work: agentic coding, long-document analysis, math, and tool use. It supports both a Thinking and a Non-Thinking mode, and tool calls work inside thinking mode. Weights live on the official Hugging Face model card under an MIT license.

Status is released, not rumored. DeepSeek V4 Pro first launched in preview on April 24, 2026, per the official DeepSeek API news post. A refined GA build, deepseek-v4-pro-0813, rolled out on August 12, 2026, which is the version most current benchmarks and pricing reference. Our earlier DeepSeek V4 release analysis tracks the naming and preview-to-GA path in detail.

DeepSeek V4 Pro at a Glance

AttributeDetail
DeveloperDeepSeek
TypeMixture-of-Experts (MoE) language model
Parameters1.6T total / 49B active per token
ModalityText in, text out (no vision confirmed)
Context window1M tokens
Max output384K tokens
ModesThinking / Non-Thinking
Pricing (input)$0.435 / 1M (cache miss); $0.003625 / 1M (cache hit)
Pricing (output)$0.87 / 1M
AvailabilityDeepSeek API (deepseek-v4-pro), DeepSeek chat, Hugging Face weights
LicenseMIT
GA builddeepseek-v4-pro-0813 (Aug 12, 2026)

How DeepSeek V4 Pro Works / What Makes It Different

The architecture is the story. DeepSeek's technical report describes a Hybrid Attention Architecture built on Token-Wise Compression plus DeepSeek Sparse Attention (DSA), designed to hold a 1M-token context at drastically reduced compute and memory cost. DeepSeek says 1M context is now the default across all its official services.

Three coined anchors matter here. Token-Wise Compression shrinks the per-token attention footprint. DeepSeek Sparse Attention (DSA) selectively attends across the long window instead of paying full quadratic cost. And the 1M Standard makes million-token context the baseline rather than a premium tier. One early hands-on note observed that "ds4pro long-context multi-point attention management feels a bit stronger than Flash," suggesting the long-context handling is a real differentiator between the two V4 models.

The Pro/Flash split is the second design axis. Flash (13B active) is the cheap, fast routing option; Pro (49B active) is the reasoning-heavy escalation path. Because both are MoE, the 1.6T total is mostly capacity, not per-token cost — only 49B is doing work on any given token.

What You Can Do With DeepSeek V4 Pro

The GA build leans hard into agentic coding. DeepSeek reported large jumps on agent benchmarks from the preview: Terminal-Bench 2.1 rose from 72.1% to 87.9%, CyberGym from 52.7% to 83.3%, and DeepSWE from 12.8% to 62.7%, with NL2Repo at 61.5%. These are vendor-reported figures.

Real-world usage reported so far clusters around a few patterns:

  • Agentic coding and automation. Developers wired deepseek-v4-pro-0813 into harnesses like OpenCode and pi, using it for asynchronous, non-real-time automation where its price makes heavy token use affordable.
  • Long-context repository work. The 1M window suits whole-codebase audits and multi-file reasoning.
  • Cybersecurity and forensics. One DeepSeek watcher noted V4 Pro "seems distinctly stronger than Flash at cybersecurity," and another flagged it doing "pixel forensics." V4-Pro is also into 'pixel forensics' https://t.co/60nfXJNMCt

Source: @teortaxesTex

  • Cost-sensitive production routing. Because output is $0.87/M, teams route bulk work here and reserve pricier frontier models for hard problems.

If you are building comparable chat or coding automation and want an API to test alongside it, the Kimi K3 chat model is a close open-frontier peer worth benchmarking against DeepSeek's V4 line.

How DeepSeek V4 Pro Compares

Positioning is "frontier-adjacent at a fraction of the price." On Artificial Analysis, the reasoning variant scored 44 on the Intelligence Index (#3 of 94 in its class) with the lowest cost-per-task ($0.04) in its comparison set. A third-party benchmark writeup placed the 0813 build only ~15 Arena points behind a leading GPT-5.6 variant on Code Arena WebDev while charging roughly 1/31st of its blended token price.

ModelInput $/1MOutput $/1MContextWeights
DeepSeek V4 Pro$0.435$0.871MOpen (MIT)
Claude Opus 4.8$5.00$25.00Not yet confirmedClosed
Kimi K3 Max$3.00$15.00Not yet confirmedOpen

The counterweight is independent evaluation. A May 2026 NIST/CAISI assessment of the preview found it the most capable PRC model tested, but placed its real capability roughly 8 months behind the U.S. frontier — closer to GPT-5 than to Opus 4.6 on held-out benchmarks. As one observer put it, "everything in the opus–fable 5 frontier seems to be within just a few points," and you likely won't notice much difference "outside one-shot demos."

Availability: How to Access DeepSeek V4 Pro

DeepSeek V4 Pro is generally available through DeepSeek's own channels. The primary paths:

  • Official API. Use base URL https://api.deepseek.com (OpenAI-compatible) or https://api.deepseek.com/anthropic (Anthropic-compatible) with model ID deepseek-v4-pro. See the DeepSeek API docs and the official pricing page.
  • DeepSeek chat. The GA build rolled out inside DeepSeek's chat product on August 12, 2026.
  • Open weights. Downloadable from the Hugging Face model card for self-hosting via vLLM, SGLang, or Transformers.

DeepSeek warns on its official page that API prices may rise, so current numbers are not guaranteed to persist.

What We Don't Know Yet

Several things remain open as of August 13, 2026:

  • GA build integrity. Some community accounts reported DeepSeek's website and platform release notices were withdrawn while the API still showed 0813, and one tester suspected a rollback. Whether this was a documentation lag, deployment error, or genuine rollback is unresolved.
  • Independent benchmark replication. The headline agent gains are DeepSeek-reported; reporting so far does not establish independent reproduction of those exact figures.
  • Quality consistency. Early hands-on reports split sharply — strong CVE rediscovery in one security test, but slow generations, mid-run cancellations, and variable 3D-rendering results in others, unconfirmed whether model or deployment.
  • Future pricing. DeepSeek flagged planned price increases without stating timing or new numbers.

Frequently Asked Questions

What is DeepSeek V4 Pro?

DeepSeek V4 Pro is an open-weights Mixture-of-Experts language model from DeepSeek with 1.6 trillion total parameters and 49 billion active per token. It supports a 1M-token context window and ships under the MIT license.

How much does DeepSeek V4 Pro cost?

DeepSeek V4 Pro costs $0.435 per 1M input tokens on a cache miss, $0.003625 per 1M input tokens on a cache hit, and $0.87 per 1M output tokens on the official DeepSeek API. Multiple community reports on August 12–13, 2026 confirmed these figures against the official pricing page.

Is DeepSeek V4 Pro open source?

DeepSeek V4 Pro ships with open weights under the MIT license, with weights hosted on Hugging Face. The training data and full pipeline are not released, so it is best described as open-weights rather than fully open-source.

What is the context window of DeepSeek V4 Pro?

DeepSeek V4 Pro has a 1M-token context window with a maximum output of 384K tokens, per DeepSeek's official API documentation. A 1M window is now the default across DeepSeek's official services.

Is DeepSeek V4 Pro multimodal?

DeepSeek V4 Pro is a text-in, text-out model. Its official specifications list only text input and output, with no image or audio modality confirmed.

DeepSeek V4 Pro vs Claude Opus?

DeepSeek V4 Pro trades close to Opus-class models on several benchmarks while costing a fraction of the price ($0.435/$0.87 versus Opus-class $5/$25). However, an independent NIST/CAISI evaluation of the preview placed it roughly 8 months behind the frontier, so it is far cheaper but not uniformly stronger.

How do I access DeepSeek V4 Pro?

DeepSeek V4 Pro is available through the official DeepSeek API with model ID deepseek-v4-pro, inside DeepSeek chat, and as open weights on Hugging Face for self-hosting.

What to Watch Next

Three signals will define whether the 0813 GA sticks. First, watch for clarification on the withdrawn release notices — a re-uploaded build or an official changelog would settle the rollback question. Second, track independent benchmark replications of the Terminal-Bench, CyberGym, and DeepSWE gains, since those are vendor-reported today. Third, monitor DeepSeek's pricing page for the flagged increase, which would change the entire cost-per-task argument.

Try related models on Kie.ai: Kimi K3, Claude Opus 4.8 and GPT-5.6.

Priya Nair

About Priya Nair

Priya covers serving costs, context windows, and the infrastructure tradeoffs behind each model launch.

View all posts by Priya Nair