The DeepSeek V4-Pro GA announcement is real now. A changelog entry dated August 13, 2026 on DeepSeek’s API docs states that “the GA release of DeepSeek-V4-Pro has been rolled out on the APP, Web, and API” — ending a strange one-day interval in which the dated version string DeepSeek-V4-Pro-0813
sat in the price list with no announcement anywhere on DeepSeek’s surfaces.
The entry itself is short, but it carries more than a status flip: a ten-metric benchmark table run on DeepSeek’s own harness, native support for the OpenAI Responses API format adapted for Codex, a three-rung thinking-effort ladder — low, high, max — that now applies to both V4-Pro and V4-Flash, and an announced pricing change that takes effect on August 16. And on the same calendar day as the changelog date, MIT-licensed weights for the -0813
checkpoint appeared on Hugging Face — roughly 893 GB of them.
This post covers the GA event itself: what the entry actually says, when it actually appeared (we dated it via the Internet Archive rather than trusting the header), what the vendor benchmark table does and does not show, what shipped to Hugging Face, what remains undisclosed, and what DeepSeek’s quiet-release choreography — price list first, announcement second — means for teams that pin model versions.
- 01GA is official — and the alias does not change.The changelog entry dated August 13, 2026 announces V4-Pro GA on app, web, and API. Callers keep using the model name deepseek-v4-pro; the pricing page's version row now reads DeepSeek-V4-Pro-0813 for the underlying checkpoint.
- 02Weights shipped the same day, MIT-licensed.The Hugging Face repo deepseek-ai/DeepSeek-V4-Pro-0813 was created August 13 with ~893 GB across 67 safetensors files. A community GGUF quant appeared the same day — the weights were genuinely downloadable, not just registered.
- 03The benchmark table is vendor-stated, vendor-run.All ten GA metrics — and every rival column on the wider Hugging Face comparison table — are DeepSeek's own runs, with its Harness in minimal mode specified for the code-agent tasks. Two rows are internal test sets nobody outside DeepSeek can rerun, and no third-party reproduction exists yet.
- 04Thinking effort is low / high / max — default unstated.Both V4-Pro and V4-Flash now expose three thinking effort levels. DeepSeek does not state which rung is the default, and “xhigh” is OpenRouter's own effort-naming layer mapped onto DeepSeek's scale — not a DeepSeek term.
- 05The release pattern rewards watching the price list.Version string in the price list first, with the Hugging Face weights repo and the changelog entry both landing inside the same ~32-hour window and the homepage copy trailing. Teams that only watch announcement feeds learned about this upgrade a day after the version label had already flipped.
01 — The AnnouncementWhat the GA entry actually says. #
The entry sits on DeepSeek’s API-docs changelog under the header “Date: 2026-08-13” and is titled “DeepSeek-V4-Pro Update.” It makes four claims, and it is worth taking them one at a time because each has a different evidentiary status.
“The GA release of DeepSeek-V4-Pro has been rolled out on the APP, Web, and API. The API calling method remains unchanged — simply set the model name to deepseek-v4-pro to use the latest version.”— DeepSeek API changelog, entry dated August 13, 2026
First, the rollout itself. GA across all three surfaces at once — consumer app, web, and API. A linked detail page adds one nuance the changelog omits: on app and web, V4-Pro is reached via an “Expert Mode” toggle (“V4 Pro is now available on app/web. Try it via ‘Expert Mode’”) rather than being the flat default. The Chinese-language changelog carries the same entry with matching content, and DeepSeek’s homepage now links to the changelog with GA copy of its own — the earlier homepage line saying the V4-Pro version “remains unchanged for now” is gone.
Second, agent capabilities. The entry’s heading reads “Significantly enhanced Agent capabilities,” with the claim that improvements are “particularly significant” in production environments, and a ten-metric benchmark table attached as evidence. We unpack that table — and its caveats — in the next two sections.
Third, two API-surface changes. The DeepSeek API now natively supports the OpenAI Responses API format, “specifically adapted for Codex” per the entry, and the thinking modes of both V4-Pro and V4-Flash gain three thinking effort levels: low, high, and max. Neither change breaks existing callers — the chat-completions path and the deepseek-v4-pro
alias keep working as before.
Fourth, pricing. The entry announces that with the official release of the V4 family, API pricing moves to a peak/off-peak structure with off-peak set at half of peak-hour prices, effective 16:00 UTC on August 16, 2026 — and the new rates on both tiers sit above today’s flat rates, a change we break down in a follow-up analysis on the new pricing windows. For this post, the load-bearing fact is simply that the change is announced, not live: at the time of writing, the flat rates still apply.
API aliasis unchanged — you call
deepseek-v4-pro
and get the GA build automatically. But the pricing page’s “MODEL VERSION” row now reads DeepSeek-V4-Pro-0813
— a dated label for the underlying checkpoint, matching the Hugging Face repo name. Stable alias on top, dated version underneath. If your compliance process assumes the model behind an alias never changes without an announcement, this release is your counterexample.## 02 — The TimelineA 32-hour window, two dates, one unchanged alias.
On August 12 we published a report documenting that the -0813 version string had appeared in DeepSeek’s price list with no announcement anywhere — no changelog entry, no news page, no Hugging Face repo, and a homepage still describing V4-Pro as unchanged. That report was accurate when it ran. This post is its planned sequel: the window it documented closed within roughly a day, and the sequence in which it closed is itself the story.
Rather than take the changelog’s “Date: 2026-08-13” header at face value, we dated the entry’s appearance through Internet Archive snapshots of the changelog page. The result bounds the publication to a roughly 32-hour window — and we deliberately will not claim a precise hour, because no snapshot exists to support one.
| When (UTC) | What appeared | How it is dated |
|---|---|---|
| By Aug 12 | Price list version row flips to DeepSeek-V4-Pro-0813; no announcement exists yet | Our Aug-12 research sweep and published report |
| Aug 12, 03:00 | Changelog still ends without the GA entry | |
One date refuses to line up neatly, and we are reporting it rather than resolving it: OpenRouter’s listing for its deepseek/deepseek-v4-pro-0813
route — described there as “the GA release of DeepSeek V4 Pro” — shows a listing date of August 12, 2026, one day before the changelog’s own header. Both dates fall inside the same Wayback-bounded window, and plausible explanations range from timezone conventions to OpenRouter provisioning the route slightly ahead of DeepSeek’s changelog publish. The honest statement is that the route and the entry appeared within the same roughly-32-hour span; anything more precise would be invention.
03 — BenchmarksTen metrics, one harness, zero third parties. #
The changelog entry attaches a ten-metric benchmark table for the GA build, and the Hugging Face model card widens it into a comparison against the April Preview, V4-Flash, and three rival models. Before any number: every cell in both tables is DeepSeek-stated. Per the model-card footnote, the code-agent tasks among the public benchmarks were evaluated with the “minimal mode of DeepSeek Harness” as the agent framework, at the max reasoning effort level with temperature 1.0 — and the rival columns are DeepSeek’s own reproductions on that same harness, not the rival vendors’ published figures. Two rows, DSBench-FullStack and DSBench-Hard, are DeepSeek’s internal test sets that nobody outside the company can rerun at all.
With that frame fixed, the cleanest reading of the table is GA against DeepSeek’s own April Preview — the same product line, the same harness, the same vendor incentive on both columns. And on that comparison the jump is large and consistent across every row.
Preview to GA · DeepSeek's own before/after, same harness
Source: DeepSeek changelog + Hugging Face model card, August 2026 — all rows vendor-run; DeepSeek's footnote specifies DeepSeek Harness, minimal mode, max effort for the code-agent tasksThe full ten-metric set from the changelog: HLE at 42.7 without tools and 60.0 with tools, Terminal-Bench 2.1 at 87.9, NL2Repo at 61.5, CyberGym at 83.3, DeepSWE at 62.7, Toolathlon-Verified at 74.1, Agents’ Last Exam at 25.7, AutomationBench (Public) at 31.8, and the two internal DSBench sets at 71.1 (FullStack) and 67.2 (Hard). Against the Preview, the gains are uneven: the narrowest public-row move is HLE without tools at +5.0 points (37.7 to 42.7), with Agents’ Last Exam next at +9.2; the largest is DeepSWE at +49.9. If those deltas survive independent reproduction once the community works through the open weights, the GA build is a materially different agent than the Preview — which is exactly what the entry’s “significantly enhanced Agent capabilities” heading claims, and exactly what nobody outside DeepSeek has yet verified.
vendor-stated: DeepSeek’s own runs on its own harness — minimal mode specified for the code-agent tasks — including every rival model’s column. At the time of writing, no independent reproduction of any GA-specific figure exists — normal for a release this fresh, and still the single most important caveat on the table. The two DSBench rows cannot be reproduced by anyone outside DeepSeek even in principle.
04 — The ComparisonCompetitive, not dominant — on DeepSeek’s own table. #
Credit where due: the Hugging Face comparison table includes rows the GA build loses, and several of them. On DeepSeek’s own numbers, V4-Pro-0813 tops the CyberGym column at 83.3 — edging the column DeepSeek labels “Fable 5 (w/ fallback)” at 83.1 — and leads AutomationBench (Public) at 31.8 over Kimi K3’s 30.8. But Kimi K3 edges it on Terminal-Bench 2.1 (88.3 versus 87.9) and beats it on DeepSWE (67.5 versus 62.7), Opus 4.8 leads NL2Repo by 8.2 points, and the Fable 5 column leads Toolathlon-Verified, DeepSWE, and HLE. On Agents’ Last Exam the GA build ties Opus 4.8 at 25.7, with Kimi K3 ahead at 27.6.
Contested rows · who leads each benchmark, per DeepSeek's own comparison
Source: DeepSeek's Hugging Face model card for V4-Pro-0813, August 2026 — every column, rivals included, run by DeepSeek on its own harnessThe honest summary: on DeepSeek’s own comparison, the GA build is a genuine step into the leading pack on agentic benchmarks — a clean sweep over its own Preview, wins on CyberGym and AutomationBench — without sweeping the field. That is a more credible shape than a table of nothing but wins, and it echoes the pattern from the V4-Flash-0731 GA release two weeks earlier, which also shipped vendor-run agent benchmarks with the same harness caveats. What it is not — yet — is evidence a procurement decision should rest on alone: the rival columns are DeepSeek’s reproductions, not the rivals’ own reported numbers, and cross-vendor tables routinely disagree by points on the same benchmark for harness reasons alone.
05 — Thinking EffortThree rungs, one unstated default. #
The most operationally useful change in the entry is the effort ladder: “the thinking modes of V4-Pro and V4-Flash now support three thinking effort levels: low / high / max.” The Hugging Face model card confirms it independently, naming the reasoning_effort
parameter and the same three values. Note what is missing from both surfaces: neither states which rung is the default. We are not going to guess — if your costs depend on it, set the level explicitly rather than relying on whatever the unset behavior turns out to be.
Effort low
The smallest thinking budget the ladder allows, on both V4-Pro and V4-Flash. DeepSeek publishes no per-rung benchmark deltas, so treat the cost/quality trade as something to measure on your own workloads.
Effort high
The intermediate step. Whether this or another rung is what you get when the parameter is unset is not stated on any DeepSeek surface we checked — an omission worth an explicit config line in production.
Effort max
The setting DeepSeek's own benchmark runs used: the model-card footnote specifies max effort for the code-agent evaluations. Published GA scores assume this rung — budget accordingly when comparing.
deepseek-v4-pro
alias says “reasoning efforts high and xhigh are supported; xhigh maps to max reasoning.” That xhighis OpenRouter’s own effort-naming layer translated onto DeepSeek’s scale — DeepSeek’s docs and model card use only low / high / max. If you see “xhigh” in a config example, you are looking at a router’s vocabulary, not the vendor’s.
06 — Open WeightsThe weights actually shipped. #
This is the part of the release that separates it from most GA announcements this month: the artifact came with the announcement. The Hugging Face repo deepseek-ai/DeepSeek-V4-Pro-0813
was created at 03:05 UTC on August 13 — the same calendar day as the changelog entry — under an MIT license, with the full weight set published: 67 safetensors files totaling roughly 893 GB in mixed fp8/fp4 precision. By 13:41 UTC the same day, a third-party GGUF re-packaging existed, which is practical evidence the weights were genuinely downloadable within hours, not merely registered.
The model card adds one architecture-adjacent line worth repeating precisely because it is so limited: the GA build “is built on the DeepSeek-V4-Pro (Preview) model structure, with a DSpark speculative decoding module attached,” and was reached via re-training rather than a structural rebuild — the same pattern the changelog described for V4-Flash-0731 in July. For what the V4 structure actually is, our V4 architecture guide and our Preview-launch coverage own that story — this post stays on the GA event.
On the GA checkpoint itself
The -0813 repo carries an MIT license — open-weight claims about the GA build are legitimate as of August 13. The older undated DeepSeek-V4-Pro repo (created in April) remains live and was not itself updated.
67 safetensors files
Mixed fp8/fp4 precision per the repo config. Sized for serious infrastructure: this is a download-and-deploy project for clusters, not workstations — and the repo ships no Jinja chat template, pointing to a custom Python encoding script instead.
From repo creation to third-party GGUF
A GGUF re-packaging of the -0813 weights appeared at 13:41 UTC the same day — about ten and a half hours after the repo was created. The open-weight release was real and immediately usable by the community.
One caution before anyone updates a spec sheet: the famous 1.6T-total / 49B-active parameter figures for V4-Pro trace only to the April Preview disclosure. The GA model card does not restate them — no DeepSeek surface we checked publishes a parameter count for the -0813 build specifically. The only place those numbers appear in this release’s orbit is OpenRouter’s own description text for the older undated alias listing. Until DeepSeek says otherwise, treat the GA build’s parameter count as formally undisclosed. Teams planning a move onto the open-weight stack should start from our V3.2-to-V4 migration playbook, which covers the operational side the model card does not.
07 — The GapsWhat the GA release does not disclose. #
A GA announcement is as notable for what it withholds as for what it states. Checked directly across the changelog, the news detail page, the pricing page, the homepage, and the Hugging Face repo and model card, the following remain undisclosed at the time of writing:
No parameter count for the GA build. As above — the 1.6T/49B figures are April Preview disclosures, restated on no DeepSeek surface for the -0813 checkpoint.No precise publish time for the entry. The changelog says only “Date: 2026-08-13”; Wayback bounds the entry’s appearance to a roughly 32-hour window (absent at 03:00 UTC on August 12, present by 11:10 UTC on August 13), and OpenRouter’s listing date reads August 12 — two dates, no resolution available.No default effort rung. Three levels are documented; which one applies when the parameter is unset is stated nowhere.No third-party benchmark verification. Every GA figure traces to DeepSeek’s own changelog, news page, or model card, all on DeepSeek’s own harness.No training-data cutoff, training-compute figure, or safety/red-teaming disclosure on any of the GA surfaces we fetched.No standard chat template. The model card is explicit that the release does not include a Jinja-format chat template — self-hosters get a custom Python encoding script instead, a real deployment friction point for anyone expecting the standard tooling path.
None of these gaps is unusual for DeepSeek, whose releases have consistently run terse. But together they define the work still open: the community now has ~893 GB of MIT-licensed weights and a vendor table to check them against, and the interesting phase of this release — independent numbers — starts where the announcement stops.
08 — The PatternPrice list first, announcement last: the quiet release. #
Step back from the individual facts and the choreography is the lesson. The dated version string surfaced in the price list first, with no announcement. The weights repo and the changelog entry both landed inside the following ~32-hour window — Wayback bounds when the entry appeared but cannot order it against the repo’s 03:05 UTC creation — and the homepage was still carrying the old copy when we last checked before that window closed. At no point did the callable alias change — anyone using deepseek-v4-pro
was moved onto the GA build without touching a line of code, and without necessarily knowing it happened.
For most teams that silence is a free upgrade. For teams that validate model behavior before adopting a new checkpoint — regulated workflows, eval-gated pipelines, anything where output drift is a defect — it is a monitoring problem: the earliest public signal of this upgrade was a version-label cell on a pricing page, not an announcement feed. The practical takeaway is to watch the surfaces that change first. Poll the pricing page’s version row, or diff the Hugging Face org’s repo list, and you learn about the next flip roughly a day before the changelog tells you. There is also a contrast worth crediting. DeepSeek’s recent history includes abrupt alias management — the July 24 alias retirement gave integrators a hard cutoff to migrate around — while this release deliberately kept the alias stable and absorbed the change underneath it. Same vendor, opposite ends of the continuity spectrum, 20 days apart. Version-pinning policy, not vendor trust, is what makes that variance survivable.
You were upgraded in place
The alias deepseek-v4-pro now serves the GA build — no code change required or possible to defer. Re-run your agent-task evals against it now: the vendor table claims large agentic gains, and your own harness is the only verification that matters.
Watch the price list, not the feed
The version row flipped a day before the changelog entry appeared. If checkpoint drift is a defect for you, monitor the pricing page's version label and the Hugging Face org directly — the announcement feed is the last surface to change.
Real weights, real friction
MIT-licensed, ~893 GB, 67 files, no Jinja chat template — deployment runs through DeepSeek's custom encoding script. The GA build's parameter count is formally undisclosed, so size hardware from the artifact, not from April's Preview figures.
Mind the August 16 switch
The GA entry announces a pricing change effective 16:00 UTC on August 16 — announced, not yet live. Model your budgets against the announced structure before the switch lands rather than after the first surprising invoice.
The forward-looking read: DeepSeek has now GA’d both ends of the V4 line 13 days apart — Flash on July 31, Pro on August 13 — each via re-training on an unchanged structure, each with same-pattern vendor benchmarks, and now with open weights landing on announcement day. If the cadence holds, the next flip will likely follow the same choreography, and the teams that treat the price list and the Hugging Face org as the real announcement channel will keep hearing the news first. For organizations deciding whether the open-weight route fits their stack at all, our AI transformation engagements start exactly there — evidence-first evaluation against your own workloads, not vendor tables.
09 — ConclusionThe sequel the price list promised. #
The announcement landed a day late — and brought the weights with it.
Three days told the whole story. On August 12, a dated version string sat in a price list with no announcement, and we reported exactly that. Within roughly 32 hours, the announcement existed: a changelog entry dated August 13 declaring V4-Pro GA on app, web, and API, a ten-metric vendor benchmark table, a three-rung effort ladder, Responses API support adapted for Codex — and, unusually for this month’s launches, ~893 GB of MIT-licensed weights on Hugging Face the same day rather than a promise of weights later.
The claims worth holding loosely are the benchmark numbers — vendor-run on DeepSeek’s own harness, rivals included, with no independent reproduction at the time of writing. The facts worth acting on are operational: your alias already serves the GA build, the effort ladder is live with an unstated default, and an announced pricing change lands on August 16. The gap worth remembering is the choreography — the price list moved before the announcement did, and it likely will again.
The open weights convert this from a claims story into a checkable one. DeepSeek has published the artifact its own table describes; the community now gets to find out how much of the Preview-to-GA jump survives someone else’s harness. That answer — not the changelog — will decide how this release is remembered.