DeepSeek V4-Pro-0813 Appears in the Price List First

DeepSeek V4-Pro-0813 exists, at the time of writing, in exactly one place DeepSeek controls: a version string inside a published pricing table. The company's API pricing page now lists the Pro column's model version as DeepSeek-V4-Pro-0813, where it previously read plain DeepSeek-V4-Pro. There is no announcement. There is no changelog entry. There are no new weights, and no benchmark for whatever sits behind the label.

What makes this worth a careful write-up rather than a headline is the gap between the artifact and the coverage. Within the same day, a marketplace listing was calling it a "GA release" and a trade outlet was reporting that the model "goes live" — while DeepSeek's own homepage banner still said, and at the time of writing still says, that V4-Pro remains unchanged for now. Two DeepSeek surfaces disagree with each other. That contradiction, not a launch, is the news.

This post documents exactly what changed on the pricing page, what did not, where every "launch" characterization actually comes from, and what teams running deepseek-v4-pro in production should do while the situation is ambiguous. Every current-state claim below was checked directly against the live primary sources on August 12, 2026.

DeepSeek's Models & Pricing page lists two models side by side: deepseek-v4-flash and deepseek-v4-pro. Each column carries two identity fields. The MODEL row names the callable API identifier — the string you actually put in a request. The MODEL VERSION row is descriptive, naming the checkpoint the alias currently serves. Flash's version row reads DeepSeek-V4-Flash-0731, matching its documented July 31 update. Pro's version row now reads DeepSeek-V4-Pro-0813.

The callable identifier did not change. The MODEL row still lists only deepseek-v4-pro, with no dated variant anywhere on the page. And the page itself is fresh: its own published-time header stamps the current revision August 12, 2026. So the observable fact is narrow and precise — DeepSeek edited a descriptive version label in its price list, and announced nothing about it on any public surface we checked.

Every outlet we found covering this story is narrating in prose from the same single artifact, and none of them published the actual diff — so here it is, field by field. The "before" column reflects the last state DeepSeek itself confirmed: the July 31 changelog entry explicitly stated that the Pro API was unchanged, and our earlier archival checks of the pricing page match it. The "now" column is the live page as directly fetched at the time of writing.

Twelve fields, two changes: a descriptive version label and one capability checkbox. The dollar figures in the Pro column are identical to what the page has carried since at least July 28, 2026. For contrast, Flash's column lists $0.0028 cache-hit, $0.14 cache-miss, and $0.28 output per 1M tokens with a concurrency limit of 2,500 — also unchanged. A rename this quiet, with zero price movement, is the kind of edit that normally precedes an announcement rather than follows one.

The substantive change hiding behind the version string is the Responses API row: Pro's cell flipped from an ✗ to a ✓. Both models now show support for JSON Output, Tool Calls, the Anthropic API format, the Responses API, and Chat Prefix Completion (Beta), with FIM Completion (Beta) marked non-thinking mode only on both.

The sequencing here is the cleanest thread in the whole story. Flash gained Responses API support as a documented, dated changelog event on July 31 — we covered that rollout and what the format convergence means in our piece on DeepSeek's Responses API and Anthropic-format support. Pro has now received the identical capability roughly two weeks later — as an undocumented checkbox flip with no changelog entry at all. Same feature, opposite evidentiary treatment. If you want the baseline this edit moved off of, the V4-Flash-0731 official release is the last Pro-relevant statement DeepSeek actually published.

The negatives here are not padding — they are the load-bearing facts, and each one is narrow and checkable. DeepSeek's official changelog still ends at July 31, 2026, twelve days before this post's publish date, with no entry for any Pro update. That July 31 entry — the Flash release — remains the newest word DeepSeek has published about Pro's status anywhere:

One caveat matters enough to state precisely. The Hugging Face model card for deepseek-ai/DeepSeek-V4-Pro still describes a preview version of the DeepSeek-V4 series with 1.6T total parameters and 49B activated — those figures describe the April Preview checkpoint, not 0813. DeepSeek has disclosed nothing about the 0813 checkpoint's parameter count, and it would be a fabrication to carry the April numbers forward. The April release that first disclosed those figures is covered in our DeepSeek V4 Preview launch analysis. And our absence claims are scoped to what we actually checked: the live pricing page in English and Chinese, the live changelog, the live homepage banner, and the Hugging Face API for deepseek-ai repositories matching the model name — all fetched directly on August 12, 2026.

Here is the detail that turns a routine catalog edit into a story. The banner on deepseek.com, fetched live the same day as the pricing page, still celebrates the Flash public beta and then states — in Chinese, "V4-Pro 暂未变动" — that "V4-Pro remains unchanged for now." Meanwhile the pricing page, stamped the same day, carries a version string that says V4-Pro has a new dated checkpoint. Both statements were live simultaneously at the time of writing.

This is worth internalizing as a pattern, not a one-off. Vendor surfaces are not versioned as a unit: pricing tables, changelogs, homepages, and docs get edited by different teams on different schedules. When surfaces disagree, the honest report is the disagreement — not whichever surface happens to support the more exciting headline.

Every "launch" characterization we found traces back to the same single artifact — the pricing-page listing — with no DeepSeek primary behind any of them. What varies is how far each outlet let its framing outrun the evidence. Laid side by side, the coverage forms a neat confidence ladder:

Two details deserve emphasis. First, the dated slug is OpenRouter's invention. DeepSeek documents only the undated alias deepseek-v4-pro as a callable identifier; nothing in DeepSeek's own docs establishes deepseek-v4-pro-0813 as a distinct model ID. OpenRouter's dated page coexists with its own undated alias page for the same model — the marketplace is tracking version strings at a finer grain than the vendor documents them, which is useful, but it means the name this story circulates under is a catalog convention, not a vendor fact.

Second, the arithmetic in the coverage is the outlets' own. Pandaily calls Pro's pricing "a 3x price premium over the Flash tier" — that is the outlet's own division of the vendor's published rates (Pro's cache-miss input and output rates each work out to roughly 3.1 times Flash's), not a ratio DeepSeek states anywhere. KuCoin's hedged item shows the same facts can be reported without the overreach. And a caution for anyone verifying this themselves: deepseek.ai is not DeepSeek's domain — the company's primaries are deepseek.com and api-docs.deepseek.com — and material hosted there attributed to leaked internal benchmarks carries no vendor authority.

The trend this illustrates is bigger than one vendor. Model catalogs, marketplaces, and aggregators now surface changes faster than vendors announce them, so a growing share of "AI news" is actually catalog archaeology — someone noticing a string change and the coverage racing ahead of the primary record. Reading those surfaces carefully is becoming a genuine operational skill, one we unpack in our companion guide to model-catalog literacy — prices, dates, and surfaces, published alongside this post.

The same pricing page carries a footnote that has been feeding a second storyline, and it needs untangling because at least one outlet has merged the two. The footnote reads, verbatim:

Three things about this footnote are checkable. It is undated and unquantified — no effective date, no percentage, no scheme; it says only that a rise is planned and that the specific plan "will be subject to official notice." It appears identically in English and Chinese — the Chinese-language pricing page mirrors it word for word. And it predates the version-string edit: our own earlier archival check of the page caught the footnote appearing in the first week of August, days before the 0813 label showed up. They are two different edits, on two different subjects, that happen to share a page.

That distinction is exactly what Pandaily's report collapses — its article treats the pricing as leaving DeepSeek room "to recoup the cost of the recent price-increase announcement," folding the earlier footnote and this week's version-string change into a single event. To be precise on both counts: DeepSeek has not announced a price increase — it has published an undated signal of intent — and no prices moved with the 0813 edit. Anyone budgeting DeepSeek spend should treat the future increase as a real but unpriced variable, not compute a number the vendor never stated.

DeepSeek has said nothing about what the 0813 label denotes, so everything in this section is inference from the company's own established conventions — labeled as such. The suffix pattern is internally consistent: Flash's official release carried -0731 for its July 31 release day, so by the same convention -0813 reads as an August 13 date — one day after this post's publish date. A pricing page carrying tomorrow's checkpoint string is the strongest hint in the whole artifact that this is staging for an announcement rather than the announcement itself.

DeepSeek's own documented precedent also bounds what a dated checkpoint can mean. The Flash-0731 changelog entry states the model "keeps the same model architecture and size as DeepSeek-V4-Flash-Preview, and was only re-post-trained" — a dated suffix, in DeepSeek's own usage, can denote a post-training-only refresh rather than a new architecture. That same entry lists nine self-reported Flash agent benchmark scores on DeepSeek's own harness (Terminal Bench 2.1 at 82.7 and Toolathlon verified at 70.3 among them, with two of the nine run on DeepSeek's own internal test sets), which the company frames as far exceeding V4-Pro-Preview. If 0813 follows the Flash playbook, the eventual announcement would bring a changelog entry, self-reported agent scores, and a re-post-trained checkpoint behind the same alias — but that is a projection from precedent, not information about 0813, for which no benchmark or specification exists at the time of writing.

The mechanics of the alias matter more than the suffix. DeepSeek's changelog tells Flash users to "simply set the model name to deepseek-v4-flash to use the latest version" — the undated alias always serves the newest checkpoint. Pro works the same way, which means whenever the checkpoint behind deepseek-v4-pro actually changes, every caller is migrated silently, with the version string on a pricing page as potentially the only public record. We covered how DeepSeek handles this alias-based versioning — and what happened when the old names were retired — in our DeepSeek API alias-retirement migration guide.

You don't need DeepSeek's announcement to act sensibly. The situation resolves into four postures depending on how you touch the model:

If DeepSeek models carry meaningful production traffic for you, this sighting is also a nudge toward cheaper insurance: alias-based vendors reward teams that keep a standing eval harness and a written model-change runbook. For the cost side of the same portfolio, our V4 Flash bulk-workload playbook covers where the cheap tier genuinely holds up. And if you want help building the eval-and-routing discipline that makes silent checkpoint swaps a non-event, that is exactly the kind of system our AI transformation engagements put in place.