Native Responses API
Tool loops can use a Responses-compatible surface instead of relying only on Chat Completions. Verify unsupported fields and tool behavior on the exact gateway you deploy.
Read this site's API contractSource-backed model guide · updated August 15, 2026
DeepSeek V4 Pro: specifications, pricing, benchmarks, and Pro vs Flash
DeepSeek V4 Pro began as the larger model in the V4 Preview family announced on April 24, 2026. DeepSeek released the V4 Pro 0813 general-availability update on August 13, 2026, with native Responses API support and Low, High, and Max reasoning effort. The published architecture still describes a 1.6-trillion-parameter mixture-of-experts model with 49 billion parameters activated per token and a one-million-token context. This page separates current product facts, earlier preview specifications, and this independent site's runtime.
1.6T / 49B
total / activated parameters
1M
published context
384K
published max output
Aug 13, 2026
V4 Pro 0813 GA
What the August 13 general-availability update changes
The main product change is agent integration, not a new marketing benchmark table. DeepSeek now documents the Responses API as native and publishes Codex setup for the same model IDs used by its API.
Current Codex path
DeepSeek's official guide covers Codex CLI, the desktop app, and the IDE extension through their shared configuration and the model IDs deepseek-v4-pro and deepseek-v4-flash.
DeepSeek Harness preview
The open-source Harness adds a plugin-based agent runtime and local Web UI. It is still a developer preview, so test it as an integration surface rather than a stable production dependency.
Open the Harness guideDeepSeek V4 Pro vs Flash specifications
Both models are mixture-of-experts systems: the total parameter count describes the complete weight set, while activated parameters describe the smaller subset used for a token. Total parameters are therefore not a direct measure of request cost or speed. The table below transcribes fields from DeepSeek’s release, model card, and API pricing page; it does not estimate unpublished values.
| Specification | V4 Pro | V4 Flash |
|---|---|---|
| API model ID | deepseek-v4-pro | deepseek-v4-flash |
| Total parameters | 1.6 trillion | 284 billion |
| Activated parameters | 49 billion | 13 billion |
| Published context length | 1 million tokens | 1 million tokens |
| Published maximum output | 384K tokens | 384K tokens |
| Reasoning effort | Low, High, Max | Low, High, Max |
| API concurrency limit | 500 | 2,500 |
| Instruct-weight precision | FP4 experts + FP8 mixed | FP4 experts + FP8 mixed |
API limits and concurrency are live-service values and can change. Recheck the official table before designing a production integration.
What changed in the V4 architecture?
Hybrid attention
The official model card describes a hybrid of Compressed Sparse Attention and Heavily Compressed Attention. The goal is to reduce long-context compute and key/value-cache demand instead of applying full dense attention uniformly across one million tokens.
mHC connections
Manifold-Constrained Hyper-Connections are presented as a change to residual signal propagation. Treat this as a model-architecture claim from the technical report, not as a feature exposed or controlled by this website.
Long-context efficiency
DeepSeek reports that Pro at a 1M context uses 27% of the single-token inference FLOPs and 10% of the KV cache of V3.2. Those are publisher measurements under its setup; they are not a promise about your provider’s latency or memory use.
For a deeper explanation of parameter counts, attention, open weights, and evidence boundaries, use the DeepSeek V4 architecture guide.
Non-think, Think High, and Think Max
V4 Pro and Flash publish three operating modes. They are not three different weight sets, and increasing reasoning effort is not automatically better for every prompt.
Non-think
Use for routine extraction, rewriting, classification, and low-risk questions where fast direct output matters. Evaluate accuracy because the model spends less explicit compute on reasoning.
Think High
Use for coding, analysis, multi-step planning, and difficult questions when extra latency is acceptable. This is usually the sensible starting point for measuring reasoning quality.
Think Max
Use for boundary testing and the hardest tasks. It can consume more output tokens and time. Compare total cost and failure rate, not only a single successful answer.
DeepSeek V4 Pro pricing explained
DeepSeek’s API bills input and output tokens separately. At the July 22, 2026 source check, V4 Pro listed $0.003625 per million cache-hit input tokens, $0.435 per million cache-miss input tokens, and $0.87 per million output tokens. Flash listed $0.0028, $0.14, and $0.28 respectively. A cache hit is cheaper, but you should not assume every repeated prompt qualifies; use provider usage records.
Example upstream API calculation
A Pro request with 100,000 cache-miss input tokens and 10,000 output tokens would use the published formula: 0.1 × $0.435 plus 0.01 × $0.87, or about $0.0522 before any other provider or application charges. This is an arithmetic example, not a quote for this website.
DeepSeek-V4.io has a separate balance system. Its Starter, Pro, and Scale packs are one-time application purchases, not the official DeepSeek API price list and not a model selector. See the pricing and cost guide for the exact free balance, pack quantities, and checkout boundary.
Publisher-reported benchmark snapshot
The values below come from the official DeepSeek V4 model card’s “comparison across modes” table. Each cell is High / Max. They compare V4 Flash and V4 Pro under the publisher’s evaluation setup; they are not independent tests and should not be combined with scores from a different benchmark version.
| Benchmark and metric | Flash High / Max | Pro High / Max |
|---|---|---|
| MMLU-Pro (EM) | 86.4 / 86.2 | 87.1 / 87.5 |
| GPQA Diamond (Pass@1) | 87.4 / 88.1 | 89.1 / 90.1 |
| LiveCodeBench (Pass@1) | 88.4 / 91.6 | 89.8 / 93.5 |
| SWE Verified (resolved) | 78.6 / 79.0 | 79.4 / 80.6 |
| MRCR 1M (MMR) | 76.9 / 78.7 | 83.3 / 83.5 |
How to use this table: treat it as a map back to the primary source. Before using a score in a decision, record the model revision, mode, benchmark release, prompts, tools, sampling, judging method, and date. For your own application, a blinded evaluation using real tasks is stronger evidence than a broad leaderboard.
When should you use Pro instead of Flash?
Start with Pro when
- The task depends on broad factual knowledge or difficult multi-domain reasoning.
- You are evaluating complex coding agents, repository-scale changes, or long tool chains.
- Long-context retrieval quality matters more than maximum request concurrency.
- A modest increase in token cost is cheaper than repeated failures or manual rework.
Start with Flash when
- You need high-throughput classification, extraction, rewriting, or simple tool use.
- Latency and concurrency dominate, and the task can be automatically checked.
- You are prototyping a workflow before spending on harder cases.
- You can route only failed or high-risk tasks to Pro after measuring a clear threshold.
A practical evaluation sequence
- Create a private set of representative easy, medium, and hard tasks.
- Define pass/fail rules, required citations, tool permissions, latency, and cost before running either model.
- Test Flash and Pro with the same context and reasoning mode, then repeat enough times to reveal variance.
- Review failures by category. Route only the categories that benefit materially from Pro.
- Re-run the set when a model alias, provider, prompt, tool, or price changes.
API, open weights, and this website are different products
Official DeepSeek API
Uses documented model IDs, provider billing, API keys, rate limits, OpenAI-format and Anthropic-format endpoints, and a live pricing table.
Read the API guideOfficial open weights
Model repositories contain exact revisions, files, precision, license, runtime guidance, and publisher evaluation data. Local deployment feasibility depends on the selected artifact.
Read the local deployment guideDeepSeek-V4.io
Is an independent application with its own account, chat, project, usage, and billing behavior. It does not imply publisher ownership or guarantee a selectable Pro runtime.
Compare product boundariesFrequently asked questions
When was DeepSeek V4 Pro released?
DeepSeek announced the V4 Preview family on April 24, 2026, then released the V4 Pro 0813 general-availability update on August 13, 2026. The preview date and the current production release are separate milestones.
Is DeepSeek V4 Pro the same as DeepSeek V4 Flash?
No. They are separate model IDs and weight sets. Pro is the larger mixture-of-experts model at 1.6T total and 49B activated parameters; Flash is 284B total and 13B activated. They share the published 1M context and API feature set, but their cost, concurrency, knowledge capacity, and difficult-task behavior differ.
Can I select Pro or Flash in this website?
Not currently. DeepSeek-V4.io uses a server-configured provider and model. The chat UI does not promise that a plan name selects V4 Pro, and the runtime model should be verified from deployment telemetry rather than inferred from branding.
Does the Pro token pack unlock DeepSeek V4 Pro?
No. Starter, Pro, and Scale are one-time balance packs for this application. They are not DeepSeek model subscriptions and do not unlock a model selector.
Is the one-million-token context always usable in one request?
One million tokens is the publisher-listed model context. Your usable request size can be lower because of provider limits, application validation, tool payloads, response reservations, memory, latency, and cost. Test the exact route you plan to use.
Are the benchmark scores independent?
No. The table on this page transcribes selected publisher-reported results from the official model card. It is useful for locating a claim, not for proving universal superiority. Reproduce tests with your own prompts, tools, settings, and review rubric.
Are the weights available for local deployment?
DeepSeek links official Pro, Flash, Base, and DSpark repositories. The exact model card and revision determine license, precision, file size, runtime support, and deployment feasibility. The availability of weights does not make a 1.6T model practical on ordinary consumer hardware.
Which source should I check first?
Use the dated DeepSeek release notice for the announcement, the official model card and technical report for architecture and evaluations, and the live API pricing table for current limits and prices. Do not copy a number from an undated comparison page.
Primary sources
- DeepSeek V4 Preview release notice — announcement date, model positioning, parameters, context, technical report, and open-weight links.
- DeepSeek change log — August 13 V4 Pro 0813 general availability, Responses API, and current agent-integration release notes.
- DeepSeek V4 Pro official model card — architecture summary, files, reasoning modes, benchmark setup, and selected results.
- DeepSeek V4 technical report — detailed methods and evaluation documentation.
- DeepSeek models and pricing — current API model IDs, limits, features, concurrency, prices, and billing rules.
Source snapshot checked August 15, 2026. Live API limits, aliases, pricing, model revisions, and product behavior can change after this date.