DeepSeek V4 Pro Release Date: Build 0813 Is Officially Live, and the Price Is Going Up

On the night of August 12–13, 2026, Beijing time, the question this page was built around stopped being a question: DeepSeek V4 Pro is released. D​eepSeek updated the deepseek-v4-pro API endpoint from the preview it had served since April 24 to the official build — 0813 — 111 days later, and Artificial Analysis re-listed the model the same way, dating it August 13, 2026 in its own FAQ. The official build keeps the preview's architecture (a 1.6-trillion-parameter mixture-of-experts with 49 billion active parameters) and keeps its price ($0.435 in / $0.87 out per million tokens), but it is a materially better agent — and it lands within a single-digit index gap of Moonshot's Kimi K3, the open-weights rival this page has been measuring it against since April.

A week ago, this page's job was to test the prediction in circulation that V4 Pro would arrive "at Kimi K3 level but 10x cheaper." The release finally lets us grade it, and the grade is closer than the preview's numbers suggested: on Artificial Analysis's independent Intelligence Index, the 0813 build scores 53 where the preview had sat at 44 — a jump that leaves the forecast's price half exactly where it was and brings its intelligence half a great deal nearer to parity. Below is what the official build actually is, what the release does to the numbers, and why the moment you have been waiting for is also the moment the vendor has told you to expect a price increase.

What the official build actually is

D​eepSeek frames 0813 as a post-training upgrade to the same architecture, not a new pretrain — the kind of release that changes the model's behaviour far more than the version number suggests. The parameter count, the 1-million-token context window and the 384,000-token output ceiling are unchanged from the preview. What changed is everything between the weights and your API call.

The two most important changes are both firsts for the model line. Native image reasoning: the preview was text-only, and the 0813 build can now analyze images, screenshots and mixed documents inside its thinking engine — the capability that had marked DeepSeek V4 Pro as a coding specialist against closed-frontier agent models. And a rebuilt agent-facing API surface: official support for the Responses API, integration into Codex-style workflows, Anthropic-format endpoints, structured JSON output and tool calls, plus two beta features aimed at agent developers — conversation-prefix continuation and fill-in-the-middle (FIM) completion.

One operational number worth knowing before you build: the official API caps DeepSeek V4 Pro at 500 concurrent requests, a fifth of the 2,500 cap on DeepSeek V4 Flash. For a chat UI that is nothing; for an agent fleet that fans out hundreds of parallel calls, it is a real ceiling — and it is the reason teams that build on this model usually route it alongside a second provider rather than bet a production path on a single connection.

Here is the changelog entry that made this page necessary in the first place — the July 31, 2026 note that announced V4 Flash's official build and closed with the sentence "The official release of DeepSeek-V4-Pro will follow soon." It did, on August 13 — quietly. There has been no formal English announcement and no English-language evaluation results from D​eepSeek as of this writing; the release is established by the API behaviour itself, the pricing page, and independent trackers, with Artificial Analysis re-listing the model as the official build the same day.

The benchmark jump, and what it does to the forecast

D​eepSeek published the preview-to-GA deltas alongside the release. They are vendor-reported figures, unreproduced by independent labs as of this writing, and they should be read as D​eepSeek's own claim about its own model — but the size of the deltas is the story regardless. The agentic and tool-use benchmarks roughly doubled:

• DeepSWE (software-engineering agent): 12.8 on the preview → 62.7 on 0813, which D​eepSeek notes surpasses Claude Opus 4.8's 58.0.

• Cybergym (AI-security agent): 52.7 → 83.3, a hair above Claude Fable 5's 83.1.

• AutomationBench: 12.8 → 31.8, ahead of Claude Fable 5's 29.1.

• Terminal-Bench 2.1: 72.1 → 87.9, within a point of Claude Fable 5's 88.0.

• NL2Repo: 38.5 → 61.5; Toolathlon: 55.9 → 74.1; DSBench-Hard: 31.1 → 67.2; DSBench-FullStack: 41.8 → 71.1; HLE: 37.7 → 42.7 (60.0 with tools).

A jump of that shape — the agentic benchmarks doubling while the architecture sits still — is the fingerprint of a heavy post-training effort, and it is exactly the shape the previous version of this article predicted. The independent checkpoints agree. Artificial Analysis has re-scored DeepSeek V4 Pro at 53 on its Intelligence Index, ranked #2 among the 104 models it currently lists — up from the preview's 44 (rank #6 of 101). On the same current snapshot, Moonshot's Kimi K3 (max) scores 60, so the intelligence gap that stood at 13 points in the preview era has narrowed to seven.

This is the moment to credit the arithmetic, because it landed almost exactly. The previous version of this article extrapolated from Flash's official build — which gained 10 index points from post-training alone — and predicted the Pro GA would land at 54. The actual re-score is 53. That is the single most important number in this update: the intelligence half of the forecast was wrong by one point, and wrong in the direction of "slightly optimistic."

One qualification belongs next to that number. The 0813 benchmark results first surfaced not in an official English post but in a WeChat group, from where they spread to Reddit and Hacker News — and the leak and the independent snapshot do not agree even directionally. On one analyst's tally, the leaked figures sit roughly 2.7% below Artificial Analysis's measurement on V4 Flash and about 9% below on V4 Pro. That is a disagreement in sign as much as in size, and it is the reason to keep filing the doubling deltas as vendor- and leak-reported until D​eepSeek posts its own evals in English and an independent lab reproduces them.

The chart above is from the original version of this page — the projection it drew (Kimi K3 at 57, "GA if it repeats Flash's +10" at 54, the preview at 44). The measured outcome sits one point below the projection: the current Artificial Analysis snapshot reads Kimi K3 at 60 and DeepSeek V4 Pro 0813 at 53. The shape of the prediction held; the residual gap is smaller than the preview suggested but still present.

For reference, here is the independent lab's snapshot of the preview era — the page this page cited in its previous version. Artificial Analysis measured DeepSeek V4 Pro (Reasoning, Max Effort) at 44, rank #6 of 101, spent $176.34 evaluating it, and logged its verbosity at 180 million output tokens against a 100-million class median. That screenshot is now historical: the same page re-listed the model at 53 after the 0813 build landed. It is also the cleanest illustration of why same-snapshot comparisons matter — the 44 and the 53 are measurements of two different builds.

What it costs — confirmed, and the fine print

The official release did not touch the price list. DeepSeek V4 Pro remains $0.435 per million input tokens on a cache miss, $0.003625 on a cache hit (a 99% discount), and $0.87 per million output tokens — about three times the price of DeepSeek V4 Flash, and still roughly a seventh of what Kimi K3 charges for input and a seventeenth for output. The cache-hit rate is effectively free, which keeps long-context workloads with repeated prefixes (system prompts, few-shot banks, codebases) dramatically cheaper than on any closed-frontier schedule.

What the release does to the pricing story is make the vendor's warning concrete. On August 6, D​eepSeek announced that it "plans to raise the overall pricing for D​eepSeek API services in the near future, with a significant increase expected," with no new rates and no effective date. That warning sat in the previous version of this page as a risk to the cheap-model thesis; it now sits next to a confirmed release, which is the natural moment for the increase to land. On top of it sits the announced-but-inactive peak-hour surcharge — 2x prices for all billing items between 09:00–12:00 and 14:00–18:00 Beijing time daily. Neither has an effective date.

The practical reading has not changed, but the clock is now visible. The $0.435/$0.87 rate is the company's published price and it is what the official build ships with; it is also, by the vendor's own written statement, a price in transition. Anyone building a cost model on DeepSeek V4 Pro today should model the current rate and the announced increase, and should know that the cheap window is open now — and that the vendor has told everyone it closes.

Because OrcaRouter passes provider list prices straight through at 0% markup, the $0.435/$0.87 on our deepseek/deepseek-v4-pro model page is D​eepSeek's own number, not a resold rate — which means the moment the announced increase lands on D​eepSeek's side, it lands here the same day, with no middleman margin and no contract cycle to wait out. For teams that want to lock in the current economics while they last — and to route around the 500-concurrency cap with automatic failover to a second provider — the model is one API call away.

The silent swap you were warned about — it happened

The previous version of this page flagged an operational hazard that matters more than the date: the API model ID is the bare string deepseek-v4-pro, with no preview suffix and no dated build string, even though the company's own changelog referred to the shipped Flash build as V4-Flash-0731. The warning was that pinning the model ID would not pin the build, and that the same string would one day silently begin serving different weights.

That day was August 13. The deepseek-v4-pro string that served the preview since April now serves the official build, with no code change required to pick it up — the exact Flash pattern the changelog described ("simply set the model name to deepseek-v4-flash to use the latest version"), repeated for Pro. If you built against the preview, you are now calling the official build, whether or not you intended to, and any evaluation you ran before August 13 is a measurement of a different model.

The practical guidance from the previous version therefore stands, one step further along: keep a regression set you can re-run on demand; keep a second model reachable without a code change; and re-run your evals now, not because the build is bad but because it is new. This is the mundane reason the failover exists — DeepSeek V4 Pro, DeepSeek V4 Flash and Kimi K3 all sit behind one OrcaRouter key on an OpenAI-compatible endpoint, so "re-run the eval on all three, then shift traffic" is a config edit rather than a procurement exercise, and a bad GA build is a routing decision instead of an incident.

Should you build on it now?

The release answers the question this page was originally written to postpone. The official build is the version to evaluate against — the preview's agent scores were a floor, and the 0813 numbers are the real ones. The honest summary of the decision: DeepSeek V4 Pro is the strongest cheap agent model with weights you can download, it is no longer "not as smart" as Kimi K3 by a wide margin — seven index points at max effort, with the vendor's own post-training curve suggesting the gap is still closing — and the vendor has told you the price is going up.

If your work sits at the frontier of hard reasoning, the seven points to Kimi K3 still buy something at the margin, and the price is a distraction from that. If your work is high-volume and merely difficult, paying roughly 14x for Kimi K3 is hard to justify when the 0813 build scores this close. And if you were waiting for the official release before building, that waiting is over — the release you were waiting for shipped, priced, with open weights and an MIT licence, on August 13, 2026.

Questions people are actually asking

Is DeepSeek V4 Pro released? Yes. The official build (0813) went live on the D​eepSeek API on August 13, 2026, replacing the preview that had served since April 24. The model string, deepseek-v4-pro, is unchanged — the build behind it is not. D​eepSeek has not yet posted an English-language announcement or its own English evals, so the release is established by the API behaviour, the pricing page, and independent re-listing rather than by a company statement.

Is it as good as Kimi K3 now? Much closer, not yet equal. On Artificial Analysis's current Intelligence Index snapshot, DeepSeek V4 Pro 0813 scores 53 against Kimi K3 (max) at 60 — down from the 13-point gap that separated the preview from Kimi K3 in early August.

Did the price change at the official release? No. The list price is unchanged at $0.435 in / $0.87 out per million tokens, with cache hits at $0.003625. The change is on the horizon, not in the list: D​eepSeek has stated it plans a "significant" increase to overall API pricing, with no rates or date yet.

Is the current price safe to build a cost model on? Conditionally, and with a visible clock. It is the company's published price for the official build, but the vendor has announced an increase is coming, and an inactive peak-hour surcharge would double daytime prices. Model both cases; even doubled, DeepSeek V4 Pro stays several times cheaper than Kimi K3.

The honest read

The forecast that started this page was half right, and the half it got right is the half people will find hardest to believe. "10x cheaper" understates it: on the same measured evaluation workload, DeepSeek V4 Pro cost $176.34 against Kimi K3's $2,437.41 — 13.8x — and the official release did not touch that arithmetic. "Kimi K3 level" is the part that came within reach but did not arrive: the 0813 build scores 53 against Kimi K3's 60 on the one composite an independent lab runs on both, where the preview had sat 13 points back. The previous version of this page predicted 54. It landed at 53.

What would change this read is specific. If the announced price increase lands with rates that halve the advantage, the cheap-model thesis stops being cheap at the hours most teams work. If a third-party lab reproduces DeepSWE 62.7 and Cybergym 83.3, the vendor's post-training jump is confirmed rather than claimed — and the WeChat-leaked eval set, which does not match Artificial Analysis even directionally, is a reason to insist on that reproduction instead of taking the deltas on faith. And if the 0813 weights ship to Hugging Face with the image and agent features intact, DeepSeek V4 Pro becomes the strongest self-hostable agent model there is — the thing that turns "cheap frontier API" into "cheap frontier infrastructure." For now, the accurate summary of DeepSeek V4 Pro is that it is released, it is the best-value serious reasoning model with weights you can download, it is closer to the frontier than it was a week ago, and its vendor has told you the price is going up.