DeepSeek V4 Pro 0813: Price, Open Weights, Benchmarks (2026)

DeepSeek shipped DeepSeek-V4-Pro-0813 on 2026-08-13, and three days later the price of running it goes up. The GA build costs exactly what the preview did. From 16:00 UTC on 2026-08-16 the whole V4 family moves to peak and off-peak rates, and every tier of that new schedule sits above today’s.

Released:        2026-08-13, GA; docs footnote first, change log entry same day
Model ID:        deepseek-v4-pro (unchanged, no 0813 suffix)
Price now:       $0.435 in / $0.87 out per 1M; $0.003625 cache hit
Price 08-16:     off-peak $0.66 / $1.98; peak $1.32 / $3.96
Open weights:    published 2026-08-13, MIT, 66 fp8 shards, ungated
Context:         1M in, 384K max output
Thinking:        on by default, effort low / high / max
Intelligence:    53 on the AA Intelligence Index, up 8 from the preview build
Cost per task:   $0.06, the cheapest of any model scoring 53
Providers:       3 (DeepSeek, plus Cloudflare and GMICloud at 3-4x)

What Is DeepSeek V4 Pro 0813?

It is the GA build of DeepSeek’s flagship, replacing the April preview behind the same model ID. The specs on the official pricing page are 1.6T total parameters with 49B active, mixture-of-experts, text only.

Nothing about the calling convention changed. You send deepseek-v4-pro and you get the 0813 build. There is no deepseek-v4-pro-0813 model name on DeepSeek’s own API, the same pattern the Flash release used two weeks earlier.

On ofox the model is deepseek/deepseek-v4-pro.

When Was DeepSeek V4 Pro 0813 Released?

2026-08-13, in two stages. The first sign was a footnote on the quick-start and pricing pages saying deepseek-v4-pro has been updated to DeepSeek-V4-Pro-0813. For most of the day that footnote was the entire announcement.

The change log entry followed later the same day, and it is the one to read: it calls this the GA release of V4 Pro across app, web and API, lists ten benchmark scores, announces the pricing change described below, and documents three thinking effort levels. It is also the follow-up to the 2026-07-31 Flash entry, the one containing the sentence “The official release of DeepSeek-V4-Pro will follow soon.”

DeepSeek’s launch post landed the same evening with one detail the docs do not carry: on the consumer app and the web client, the GA build is not the default. It sits behind Expert Mode. If you are comparing app answers against API answers, that toggle is the difference between testing 0813 and testing whatever the app serves by default.

The practical lesson is about polling rather than about DeepSeek. A vendor’s model can be live in the API hours before the page you monitor says anything, so a release check that reads only the change log will be late, and one that reads only the model list will miss the pricing note attached to it.

Worth knowing if you script against those docs: the /news/ paths are a catch-all. Requesting a news URL that does not exist returns HTTP 200 with the quick-start page body, so a status-code check will tell you a nonexistent announcement is live.

Is DeepSeek V4 Pro 0813 Open Source?

Yes, as of 2026-08-13, and the timing is worth knowing because most of the release-day coverage was written before it was true.

• deepseek-ai/DeepSeek-V4-Pro-0813 holds 66 safetensors shards in fp8, MIT, ungated, alongside a config.json and DeepSeek’s own encoding utilities.
• The repository is stamped 2026-08-13, but it did not appear in the organisation’s public model list when we checked that morning UTC. It was there by the evening, roughly half a day after the API went GA.
• The older deepseek-ai/DeepSeek-V4-Pro repo, last modified 2026-06-22, is the April preview: 64 shards, also MIT. Do not grab that one by accident; the names differ by a suffix and the builds differ by eight index points.
• unsloth/DeepSeek-V4-Pro-0813-GGUF had already staked out the name earlier in the day with nothing in it but a README. Quantisations are what fills in next.

Two things follow. Artificial Analysis had tagged the 0813 entry “Open weights model” while no such repository was reachable — the label was early rather than wrong, the same way it was early on Flash 0731. And any comparison table that lists V4 Pro 0813 as closed-weight, including several written on release day, is now describing a window that lasted less than a day.

How Much Does DeepSeek V4 Pro Cost?

$0.435 per million input on a cache miss and $0.87 output until 2026-08-16, then peak and off-peak rates that are higher in both halves.

The cache hit rate is what actually decides your bill here, since a hit costs 120x less than a miss. We worked through that math on the preview build in DeepSeek V4 Pro real cost, and the ratios are unchanged.

What Changes on 2026-08-16?

The vague warning DeepSeek posted on 2026-08-06 — “a significant increase expected”, no amount, no date — became a schedule with the GA release. From 16:00 UTC on 2026-08-16, billing splits into peak and off-peak, with off-peak set at half of peak. Peak hours are 01:00–04:00 and 06:00–10:00 UTC; everything else is off-peak.

Three things to take from that table. Off-peak is not a discount against today’s rate, it is a smaller increase: V4 Pro output goes up 2.3x off-peak and 4.6x at peak. The cache-hit line moves most in relative terms, 6x off-peak, which quietly weakens the main lever people use to keep DeepSeek cheap. And the two peak windows are short and awkwardly placed — 01:00–04:00 and 06:00–10:00 UTC covers most of the Chinese working day, so whether this reads as a rise or a small rise depends mostly on where your batch jobs run.

The eight-point jump and the price rise arrived in the same change log entry, which is the honest way to read this release: DeepSeek is charging for the improvement rather than absorbing it. For the levers that still work after the 16th, see six ways to pay less before the increase.

What Is the Context Window and Max Output?

1M tokens in, 384K maximum output, the same on both V4 models.

That output ceiling is unusually high. Most flagships cap output between 64K and 128K, so a single V4 Pro call can legitimately produce a book-length response. It also means max_tokens is not a safety net by default, and a runaway generation on Pro can bill 384,000 output tokens at $0.87 per million before it stops.

For how the window is actually consumed by system prompts, tool schemas and reasoning tokens, see what is a context window.

How Does V4 Pro 0813 Score on Benchmarks?

Two sets of numbers exist, and they are not interchangeable: DeepSeek’s own agentic scores, and Artificial Analysis’s independent index, where V4 Pro 0813 reads 53, up from 45 on the preview build.

DeepSeek’s change log entry publishes these for the GA build. They are vendor-run and there is no methodology note attached — worth stating plainly, because the Flash 0731 entry six weeks earlier did specify its harness (DeepSeek Harness minimal mode, max effort, top_p 0.95, temperature 1.0) and this one specifies nothing.

Both columns come from DeepSeek’s own change log, so the comparison between them is internally consistent even though neither is independently reproduced. Two of the ten are internal test sets DeepSeek does not release — DSBench-FullStack and DSBench-Hard — so those rows cannot be checked by anyone outside the company at all.

Artificial Analysis is the independent read. Where V4 Pro 0813 sits on the full leaderboard, Index v4.1.1, snapshot 2026-08-13:

The row that matters is the cost column, not the rank. Three models tie at 53, and V4 Pro 0813 runs about 5x cheaper per task than either of the other two.

Two cautions on how this number gets quoted:

• AA’s model page says “#2 / 104” and that is not an overall rank. It is the position within open weights models of the same size class. On the full board, 53 puts it around 17th.
• The index is a rolling benchmark. Flash 0731 read 50 when it launched and reads 52 now, because AA moved from v4.1 to v4.1.1 and recomputed. Cite the ordering rather than the absolute number, and date whatever you cite.

Three supporting figures from the same snapshot, all measured on the max-effort configuration:

• Verbosity 128M output tokens across the index run, against 206M for Flash 0731. Pro reaches a higher score using fewer words, which is the opposite of the usual pattern.
• 83.2 output tokens per second, first chunk at 1.63s. For comparison, Claude Opus 5 at max effort takes 63.9s to first chunk and GPT-5.6 Sol at max takes 176.6s.
• $135.03 total to run the full index, versus $72.03 for Flash 0731.

Where Can You Run V4 Pro 0813 Besides DeepSeek?

Two other places so far, both at three to four times the price, and the list is growing by the day. The OpenRouter endpoints API listed two providers for deepseek/deepseek-v4-pro-0813 on 2026-08-13 and three on 2026-08-14:

Flash 0731 has 28 listed endpoints, and that is the shape Pro is heading towards now that its weights are public — the third provider appeared within a day of the repository going up. Two practical consequences, both true today and both with a shelf life:

• No cheaper third-party route exists yet. With Flash, shopping hosts is a real lever, and cache-read rates alone vary about 27x between them. Every alternative Pro host currently costs 3-4x the source rate, and GMICloud’s cache reads run $0.145 against DeepSeek’s $0.003625.
• Failover exists, but budget for it. If DeepSeek’s endpoint is saturated you can retry elsewhere at three to four times the bill and without the implicit prompt caching DeepSeek’s own endpoint advertises. That is a fallback for an outage, not for steady traffic — at least until enough hosts show up to compete on price.

Watch the slug when you wire this up. deepseek/deepseek-v4-pro-0813 is the new build. Plain deepseek/deepseek-v4-pro on OpenRouter is still the April preview and carries a summary price of $1.168 in / $2.336 out, roughly 2.7x the current official rate for a weaker model. The Flash release had the identical trap.

How Do I Access DeepSeek V4 Pro 0813?

Point an OpenAI-compatible client at DeepSeek and send deepseek-v4-pro. Both the OpenAI and Anthropic protocols work, and as of this release the Responses API works on Pro too.

from openai import OpenAI

client = OpenAI(
    base_url="https://api.deepseek.com",
    api_key="YOUR_DEEPSEEK_API_KEY",
)

response = client.chat.completions.create(
    model="deepseek-v4-pro",
    messages=[{"role": "user", "content": "Hello!"}],
    reasoning_effort="high",
)
print(response.choices[0].message.content)

The Responses API support is new for Pro. On 2026-07-31 DeepSeek’s position was that native Responses support shipped with Flash and Pro would follow in early August. The pricing page now marks it supported on both, and the GA change log entry says the format is “specifically adapted for Codex” with a one-click configuration script. That matters mainly for Codex CLI, since that client only speaks wire_api = "responses".

Why Does Thinking Behave Differently Depending on Where You Call It?

Reasoning defaults are not part of a model. They are a property of whatever endpoint you happen to be calling, and nothing in a response tells you which default you got.

DeepSeek’s own thinking mode documentation is explicit that thinking is on by default for V4 with effort set to high, and as of this release it documents three effort levels rather than one switch: low, high and max. It also publishes the mapping for the OpenAI-style values that have no equivalent — medium and xhigh both resolve to high — so two of the five strings you might send are silently collapsed even on DeepSeek’s own endpoint.

Route the same model ID through a gateway and the default itself can change, without any error and without any field in the response marking it. We measured this on 2026-08-13 against deepseek/deepseek-v4-pro on an OpenAI-compatible gateway with one bat-and-ball prompt, three runs per configuration:

Three things fall out of that table. Thinking was off by default on this route, against the vendor default of on, and consistently so. Turning it on cost fifty to seventy-five times the output tokens for the same one-word answer, which is the entire billing story on a reasoning model — and note the spread within each row, which is why a single measurement of reasoning cost is worth very little. And DeepSeek’s native thinking parameter did nothing: its three runs land in the same range as sending no parameters at all, while OpenAI’s reasoning_effort moves the number every time. The parameter name you need depends on the gateway, not on the model.

Reasoning tokens were billed as completion tokens in every case without the reasoning text being returned. That is normal, and it is also why a bill can triple while the visible output stays the same length.

The check that costs nothing: send one arithmetic prompt with no reasoning parameters and read usage.completion_tokens, three times rather than once. Single digits every time means thinking is off. Hundreds means it is on and you are paying for it. Do this against any new endpoint before you size a budget on it, and repeat it when a provider silently updates a model behind a stable ID, which is exactly what happened here.

If you want to compare two models’ reasoning economics without opening two vendor accounts, an OpenAI-compatible gateway lets you swap the model string and keep everything else fixed. On ofox the two IDs are deepseek/deepseek-v4-pro and deepseek/deepseek-v4-flash, against the same base_url and the same key.

for model_id in ["deepseek/deepseek-v4-pro", "deepseek/deepseek-v4-flash"]:
    r = client.chat.completions.create(
        model=model_id,
        messages=[{"role": "user", "content": PROMPT}],
        reasoning_effort="high",
    )
    print(model_id, r.usage.completion_tokens)

One caveat on that route, in the interest of not overclaiming: the ofox catalog still stamps deepseek/deepseek-v4-pro with a 2026-04-24 release date and its model page still shows a 2026-07-12 LMArena snapshot, so you cannot confirm from the outside which build is behind it today. The parameter behaviour above is independent of that question.

DeepSeek V4 Pro 0813 vs V4 Flash 0731

One index point apart, double the price per task.

Pro closed most of the quality gap this release and none of the price gap. Flash remains the default for agent loops and high-volume work, on cost, concurrency and failover options all at once. Pro earns its premium on the hardest single calls, where one point of index score is worth twice the money.

The full task-type breakdown is in V4 Pro vs Flash, which was written against the preview build and the routing guidance still holds, now by a narrower margin.

References

• DeepSeek API pricing
• DeepSeek first API call
• DeepSeek change log
• DeepSeek launch announcement, 2026-08-13
• DeepSeek thinking mode documentation
• Hugging Face: deepseek-ai/DeepSeek-V4-Pro-0813
• Hugging Face: deepseek-ai/DeepSeek-V4-Pro
• Hugging Face: unsloth/DeepSeek-V4-Pro-0813-GGUF
• Artificial Analysis: DeepSeek V4 Pro 0813
• Artificial Analysis model leaderboard
• OpenRouter: deepseek/deepseek-v4-pro-0813