DeepSeek V4 Pro 0813: Price and Benchmarks

Model launches have settled into a ritual: a teaser video, a wall of benchmark charts, an excited founder post. DeepSeek skipped all of it. DeepSeek V4 Pro 0813 went generally available on August 13, 2026 with no announcement at all. One line landed in the API changelog and the weights appeared on Hugging Face.

A quiet launch does not mean a small one. This is a 1.7 trillion parameter model, MIT licensed and downloadable, served through an API at roughly one thirtieth the output price of the closed frontier.

Below we look at both the company’s own numbers and independent measurements, because the two do not tell the same story. 👇🏻

What Is DeepSeek V4 Pro?

DeepSeek V4 Pro is the large model in DeepSeek’s V4 family. The family has two members: the fast and cheap V4 Flash at 304 billion parameters, and V4 Pro. The 0813 suffix is the release date of this revision, exactly like the 0731 on Flash.

Here is the spec sheet:

Two things stand out.

First, the license. MIT is about as permissive as it gets: download it, change it, put it in your commercial product, report to nobody. Coming three days after Meta shipped Muse Glimmer under Apache 2.0, it shows where the open-weight race is heading.

Second, the context window. One million tokens is roughly 750,000 words: an entire mid-sized codebase, or a few thousand pages of documentation, in a single request.

Curious how many tokens your own text takes? Paste it into our token counter and see it instantly.

What Changed in This Release?

The company shipped three things.

Three reasoning efforts. The model now takes low, high or max, controlling how much it deliberates before answering. On simple classification work low is both faster and cheaper; on multi-step agent work max thinks deeper. V4 Flash got the same control.

DSpark speculative decoding. A speedup module is now attached to the model. Speculative decoding works like this: a small, fast draft model guesses the next few tokens, and the large model verifies all of those guesses in one pass. When the guesses land, you get several tokens for the price of one step. Quality is unchanged, only throughput moves. It is enabled behind a flag in both vLLM and SGLang.

Responses API and Codex support. DeepSeek now speaks OpenAI’s Responses API format natively. In practice that means tools and agent frameworks written for OpenAI can point at DeepSeek by changing a base URL and a model name. Codex integration comes with it.

Official Benchmark Results 📊

The table published on the model card puts the new revision next to both its own family and its competitors:

The real story in this table is not the competitor columns, it is the preview column. DeepSWE went from 12.8 to 62.7, Cybergym from 52.7 to 83.3, Terminal Bench from 72.1 to 87.9. The GA model is not the preview with a few patches on top; on agentic work it behaves like a different model.

Against the competition the picture is balanced: first place on Cybergym and AutomationBench, behind Kimi K3 and Fable 5 on Terminal Bench and DeepSWE, and roughly level with Opus 4.8 across most rows.

But a warning belongs here.

What Does Independent Testing Say? ⚖️

Artificial Analysis ran the model through its own Intelligence Index and landed on 53, clearly behind the closed frontier. Their per-benchmark numbers:

Look at Terminal Bench 2.1 carefully. The company reports 87.9, the independent lab reports 79. That nine point gap is not an accident. Vendors run their own evals with tuned settings, at the highest reasoning effort, sometimes inside a custom scaffold. Independent labs use one standard harness for every model.

Two more numbers from the independent run: the model produces 83 tokens per second and costs roughly $0.06 per task on the index. That cost sits far below the closed frontier.

The real question: Pro or Flash?

Here is the sharpest detail in this release. On the Artificial Analysis index, V4 Pro scores 53 and V4 Flash scores 52. One point apart.

The price gap is more than three times. Judged on the index average alone, the money you spend upgrading to Pro buys you nothing.

The difference shows up in specific work rather than in the average: 62.7 against 54.4 on DeepSWE, 60.0 against 51.5 on HLE with tools. Long agent chains and tool-using tasks are where Pro pulls clear. For chat, summarisation, classification and short code completion, the one third price of Flash is the obvious call.

Pricing and the August 16 Change 💸

Per million tokens:

For comparison, Claude Opus 5 costs $5 in and $25 out per million tokens, and Grok 4.6 costs $2 and $6. On output, DeepSeek V4 Pro is roughly one thirtieth of Opus.

Pay attention to the cache row. Cached input drops to $0.003625, about one percent of the normal rate. If you are building an agent that resends the same system prompt or the same source file over and over, that is where the savings actually live.

From August 16, 2026 at 16:00 UTC a new tariff structure kicks in: peak and off-peak hours are separated, and off-peak rates are set at half the peak price. Moving batch work such as data labelling, archive summarisation or bulk translation into those windows cuts the bill in half outright.

What does that mean for your workload? Put your monthly input and output token counts into our LLM cost calculator and compare DeepSeek against Claude and GPT side by side.

Open Weights: Can You Actually Download It?

Yes, but be realistic.

The weights sit in the deepseek-ai/DeepSeek-V4-Pro-0813 repository on Hugging Face under the MIT license, in mixed BF16 and FP8 tensors. It runs under vLLM and SGLang, with DSpark acceleration behind a separate flag.

The problem is size. 1.7 trillion parameters at FP8 precision still means roughly 1.7 terabytes on disk and total VRAM in that same range. That does not fit in a single eight-GPU server; it needs a multi-node cluster. So “open weights” here does not mean you can run it at home. It means you can run it on your own infrastructure with no dependency on anyone, which is exactly the distinction that matters to organisations that cannot send their data outside.

If running something locally is the actual goal, even the 304B Flash is too big. Meta’s 30B Muse Glimmer, which fits on a 24 GB consumer card, is the realistic reference point.

For wiring open models into coding tools, our guide on using Claude Code with Ollama and open models covers the practical setup.

For Developers: Five Lines to Start

The API is compatible with the OpenAI library. Code written for ChatGPT connects to DeepSeek with two changed lines:

from openai import OpenAI

client = OpenAI(
    api_key="YOUR_KEY",
    base_url="https://api.deepseek.com",   # this is the only difference
)

response = client.chat.completions.create(
    model="deepseek-v4-pro",
    messages=[{"role": "user", "content": "Explain DeepSeek V4 Pro in one sentence."}],
)

print(response.choices[0].message.content)

Three things worth knowing:

• The model name is not pinned. Asking for deepseek-v4-pro always routes to the latest revision, currently 0813. Use the full name if you need to pin it.
• Use the cache. Repeated system prompts bill at one percent of the normal input rate. On long-context agents this is where most of the bill disappears.
• Pick the effort per task. low, high, max. Do not run everything at the default.

For local serving the recommended settings are temperature=1.0, with top_p=0.95 for agentic work and top_p=1.0 for standard use.

Who Is This For?

Move to V4 Pro if you are running long agent chains, terminal or repository-wide coding tasks, genuinely filling a million token context, or you have to keep data on your own infrastructure.

Stay on V4 Flash if your work is chat, summarisation, classification, translation or short code completion. One point of index difference, three times the price.

Stay on a closed model if you are doing the hardest reasoning work. Gaps like 18 against 32 on CritPt and 49 against 60 on SciCode show a real remaining deficit on deep scientific reasoning.

Frequently Asked Questions

Q: When was DeepSeek V4 Pro released? A: August 13, 2026. DeepSeek published no announcement; it added one line to the API changelog and uploaded the weights to Hugging Face.

Q: How much does DeepSeek V4 Pro cost? A: $0.435 per million input tokens and $0.87 per million output tokens. Cached input drops to $0.003625. From August 16, 2026 off-peak hours are billed at half rate.

Q: Is DeepSeek V4 Pro free? A: API usage is paid, but the weights are free to download under the MIT license. The DeepSeek app and web interface also offer limited free use.

Q: Is DeepSeek V4 Pro open source? A: The weights are published on Hugging Face under the MIT license with essentially no restrictions, commercial use included. Training data was not released, so the accurate term is open weights rather than open source.

Q: How many parameters does DeepSeek V4 Pro have? A: Roughly 1.7 trillion. It uses a mixture-of-experts architecture, so only a subset of those parameters is active for any given token.

Q: What is the difference between V4 Pro and V4 Flash? A: Pro is ~1.7T parameters, Flash is 304B. Pro leads clearly on agentic and coding benchmarks (62.7 vs 54.4 on DeepSWE, 60.0 vs 51.5 on HLE with tools), but the Artificial Analysis index puts them one point apart at 53 and 52. Pro costs more than three times as much.

Q: Is DeepSeek V4 Pro better than Claude Opus 5? A: Not in independent testing. It scores 79 against 89 on Terminal-Bench v2.1 and 55 against 67 on GDPval-AA v2. Its output tokens cost about one thirtieth as much, so the comparison belongs on price-performance rather than raw capability.

Q: Can I run DeepSeek V4 Pro on my own computer? A: In practice, no. 1.7 trillion parameters at FP8 still requires terabyte-scale total VRAM and a multi-node cluster. For home hardware, look at models in the 30 billion parameter range.

Take care… 🙂