DeepSeek shipped V4-Pro with no announcement: the only official evidence is a version string and a price table

DeepSeek released the production build of its flagship model, DeepSeek-V4-Pro-0813, in August 2026 by swapping a version string in its API documentation rather than publishing an announcement. The official DeepSeek API pricing page lists the deepseek-v4-pro endpoint as DeepSeek-V4-Pro-0813 and documents $0.435 per million input tokens on a cache miss, $0.003625 on a cache hit, and $0.87 per million output tokens, yet the same documentation's changelog contains no August entry at all and the deepseek-ai organization on Hugging Face has no repository for this build. The benchmark table now widely quoted for this model originated in the official DeepSeek WeChat group, passed through a Reddit post that was later deleted, and surfaced on Hacker News, and it appears in no official DeepSeek channel. ASAP separates what the documentation confirms from what it does not, and works out what this release actually changes.

The only trace DeepSeek left is one changed row in a documentation table

DeepSeek acknowledges DeepSeek-V4-Pro-0813 in exactly one place of its own: the version column of the API pricing table. That table lists deepseek-v4-pro as DeepSeek-V4-Pro-0813 and deepseek-v4-flash as DeepSeek-V4-Flash-0731. Both models carry a 1M-token context and a 384K-token maximum output, both support thinking mode, JSON output and tool calls, and both accept requests in either the OpenAI format or the Anthropic format, the latter at the base URL api.deepseek.com/anthropic.

None of the material that normally accompanies a production release is present. The most recent entry in DeepSeek's API changelog is the V4-Flash update dated July 31, 2026, with no August entry. The most recently updated model in the deepseek-ai organization on Hugging Face is DeepSeek-V4-Flash-0731, and no repository named DeepSeek-V4-Pro-0813 appears in the listing. The same organization still hosts DeepSeek-V4-Pro, updated June 22, and DeepSeek-V4-Pro-DSpark, updated July 4, which means this production build alone sits outside the weight releases.

Developer Simon Willison wrote on August 12, 2026 that the model had no obvious official announcement page, and linked the OpenRouter model page instead. That page describes a mixture-of-experts system with 1.6 trillion total parameters and roughly 49 billion activated per token, supporting high and xhigh reasoning effort levels along with tool calling and structured outputs. The model reached production traffic with its distribution path documented and its performance evidence absent.

The price table is unchanged, and a warning of a steep increase sits on the same page

DeepSeek's pricing page carries both the current rates and an explicit warning of an across-the-board increase. The documentation states that the company plans to raise overall pricing for DeepSeek API services in the near future, that a significant increase is expected, and that users should plan accordingly. No effective date and no magnitude are given. Current rates are $0.435 per million input tokens on a cache miss, $0.003625 on a cache hit and $0.87 for output on deepseek-v4-pro, against $0.14, $0.0028 and $0.28 on deepseek-v4-flash.

Dividing the two tiers exposes the design of this table. On cache-miss input and on output, Pro costs 3.1 times Flash, the identical multiple in both rows, but on cache-hit input it costs only 1.3 times as much. Moving up a tier is expensive for tokens the model has never seen, and nearly tier-neutral for context already sitting in cache. Inside Pro alone, the cache-miss rate is exactly 120 times the cache-hit rate. That ratio was 50 on Flash, so the higher tier more than doubles the degree to which cache reuse decides the bill.

Overlaying the price warning on that structure changes the order of an adoption decision. Fixing an annual budget on the currently published rates contradicts what the documentation itself warns. With neither the magnitude nor the affected rows known, the most defensible design is the one that raises the cache-hit ratio. A pipeline that holds a fixed prefix and varies only the query keeps most of its bill on the cache-hit rate and absorbs the shock, while a pipeline that refills its context on every call takes the increase in full.

A 500 against 2,500 concurrency limit reveals the intended deployment shape

DeepSeek documents a concurrent request limit of 500 for deepseek-v4-pro against 2,500 for deepseek-v4-flash, a factor of five. The tier that costs 3.1 times more is capped at one-fifth of the throughput. This number appears in no benchmark table, yet it binds a service design earlier than any score does.

Two things follow. The first concerns capacity: a 1.6-trillion-parameter model structurally serves fewer simultaneous requests from the same cluster, and a 500-request ceiling indicates DeepSeek is not positioning this tier for bulk traffic. The second concerns deployment shape: a product that calls the top tier on every user request hits this ceiling first, while overnight batches and selective routing of hard tasks fit inside it comfortably. That points the same direction as the pricing structure. The documented constraints consistently describe a model meant for selective calls rather than standing ones.

Why ASAP does not reproduce the circulating benchmark table

No performance figures for this production build have been posted in any official DeepSeek channel. By Willison's account, the distribution chain ran from a table posted in the official DeepSeek WeChat group, to a copy in a Reddit post that was subsequently deleted, to an ASCII rendering on Hacker News. Every figure quoted since has passed through that chain, and no public address holds the original table.

The distinction that matters is the evidentiary grade of the preview figures against the production ones. The preview scores are officially sourced: the comparison table in the DeepSeek-V4-Flash-0731 model card published on July 31 includes the V4-Pro preview at 72.1 on Terminal Bench 2.1, 52.7 on Cybergym, 12.8 on DeepSWE, 38.5 on NL2Repo and 55.9 on Toolathlon-Verified. The numbers circulating for the 0813 build carry nothing of that grade. Placing both rows in a single table because the benchmark names match would make a verified value and an unverified one look visually equivalent, which is why ASAP does not redraw it.

The episode also reads as a signal about release practice itself. For two years the pattern for a major model release has been a model card, an evaluation table and a license notice shipped together, and DeepSeek followed exactly that pattern for Flash on July 31. Two weeks later the same company shipped a higher-tier production build through a documentation version bump, which means published documentation is no longer a required step in its release process. As the basis for judgment moves from official documents to private group chats, the set of people able to verify a number narrows to whoever received the original.

The table where the lower tier beat the higher tier has not been redrawn

The tier inversion created on July 31 inside DeepSeek's own lineup remains officially unresolved after this release. The Flash-0731 model card reported, in DeepSeek's own table, that the lower-tier Flash led the higher-tier V4-Pro preview on all nine benchmarks, and because the comparison ran against a preview it was explicitly a table awaiting a redraw once the production Pro shipped. The production Pro has shipped, and the official table has not.

DeepSeek's API therefore now exposes two tiers priced 3.1 times apart with no official basis on which to compare their performance. The last official document a customer can consult still shows Flash ahead of the Pro preview. With no documented case for the higher tier, choosing Pro today rests either on figures from an unofficial channel or on measuring the two directly.

Only the second is defensible in practice. Both tiers are exposed through the same request formats and their cache-hit rates differ by only 1.3 times, so pushing a 100-task internal evaluation set through both endpoints with identical prompts finishes within a day at modest cost. Filling the gap left by the missing public benchmark with an internal evaluation is the practical task this release hands to engineering teams.

What a production build without open weights means for regulated adoption

The absence of weights on Hugging Face is a harder constraint than licensing for organizations handling regulated data. DeepSeek-V4-Flash-0731 shipped as MIT-licensed open weights and can be self-hosted, while DeepSeek-V4-Pro-0813 is absent from the organization listing and leaves no option other than API calls. For finance, public sector and healthcare workloads that cannot send data across a border, that difference is not a choice of tier but a question of whether adoption is possible at all.

Existing repositories make the situation easier to misread. The deepseek-ai organization still lists DeepSeek-V4-Pro, updated June 22, and DeepSeek-V4-Pro-DSpark, updated July 4, so anyone arriving through search may assume the production weights are already available. In fact the build the API points to and the builds in the repository are different. Teams evaluating self-hosting should verify what they are downloading against the version string in the API documentation rather than the repository name.

The configuration available under these constraints is a split by tier: route work that may leave the perimeter to the Pro API, and handle data-resident work on a self-hosted Flash-0731 under its MIT license. The two share request formats and the same 1M-token context, so a routing layer keeps switching costs low. The premise of that design is an internal evaluation showing Flash is sufficient for the work in question, and as above, that evaluation has to be run rather than read.

What will settle this

Three developments will determine how this release is read. The first is whether official documentation follows. If DeepSeek adds an August changelog entry and publishes an evaluation table for 0813, the status of the circulating figures resolves at once. If it moves to the next build without doing so, its models become comparable only through third-party measurement.

The second is a weight release. DeepSeek-V4-Pro from June 22 and DeepSeek-V4-Pro-DSpark from July 4 are already in the organization, so publishing 0813 is a policy choice rather than a technical one. If keeping only the top tier closed hardens into a pattern, DeepSeek's open-weight strategy should be read as scoped to its lower tier.

The third is the terms of the announced price increase. The documentation warns of a significant rise without naming a date or the rows affected, so whether the 120-fold gap between the $0.003625 cache-hit rate and the $0.435 cache-miss rate survives will govern real operating costs. If that ratio narrows, the economics of every pipeline designed around fixed-context reuse move with it.

Source: DeepSeek official API documentation pricing page and changelog (api-docs.deepseek.com), the deepseek-ai organization model listing on Hugging Face, the OpenRouter deepseek/deepseek-v4-pro model page, Simon Willison's post of August 12, 2026, and the comparison table in the official DeepSeek-V4-Flash-0731 model card, compiled by ASAP.