DeepSeek V4 — the V4-Pro and V4-Flash pair that has led the price-performance charts all spring — has technically been a preview since it shipped on April 24, 2026. As of August 1, the saga has its first ending: V4-Flash's stable build (0731) shipped on July 31 — same architecture, fully retrained, dramatically stronger on agent work, weights on Hugging Face the same day — while V4-Pro's stable release is dated early August, with a Pro cache-price change scheduled for August 3 that the community reads as the likely day. The mid-July window slipped, the release split in two, and on the evidence the Flash half was worth the wait: full launch breakdown here. What did not slip: the legacy deepseek-chat and deepseek-reasoner aliases stopped working on July 24, 2026, 15:59 UTC — the deadline held even though the launch did not. If something in your stack just went dark, that is why; the fix is a one-string rename to deepseek-v4-flash or deepseek-v4-pro.
Confirmed, reported, rumored
Confirmed (DeepSeek's own documentation):
- V4 Preview shipped April 24, 2026: V4-Pro (1.6T total parameters, 49B active) and V4-Flash (284B total, 13B active), both MIT-licensed open weights with a 1M-token context, up to 384K output tokens, and dual thinking / non-thinking modes.
- The July 24 deadline: the legacy
deepseek-chatanddeepseek-reasoneraliases are retired and become inaccessible after July 24, 2026, 15:59 UTC. Anything still pointed at them breaks that day. - Current official API prices stand at $0.435 input / $0.87 output per million tokens for V4-Pro, and $0.14 / $0.28 (cache hits $0.014) for V4-Flash.
Reported (industry press, not yet in DeepSeek's docs):
- A mid-July stable release for V4, replacing the preview label.
- The interesting part: time-of-day pricing on DeepSeek's own API. Reported peak windows are Beijing time 9:00–12:00 and 14:00–18:00, where prices double; outside those windows, today's prices hold.
Rumored (community sightings, unverifiable):
- A viral video claims a gray-test build of the stable V4 generating a playable "Minecraft plus No Man's Sky" game from one prompt — read as evidence the stable build is stronger than the preview. Cherry-picked single prompts prove little; the skeptics' counterexample in the same threads is apt: Gemini 3.1 Pro shipped in February and is still labeled preview. Gray-test sightings are not a release.
- Timing speculation clusters around WAIC (the week of July 17), the same window the Kimi K3 signals point to — see our K3 tracker.
July 31: Flash ships stable — the saga's first ending
The release finally began, and it began at the bottom of the price ladder. DeepSeek-V4-Flash-0731 entered official public beta on July 31: identical 284B/13B architecture, "only re-post-trained" — and the retrain posted 82.7 on Terminal-Bench 2.1 and 25.2 on Agents' Last Exam against Opus 4.8's 25.7, beating DeepSeek's own V4-Pro preview across nine agent benchmarks (vendor numbers, pending reruns). Prices unchanged at $0.14/$0.28. Weights hit Hugging Face the same day. Full analysis: the launch breakdown.
V4-Pro stable: early August — August 3 passed without it, and the signals have stacked up since. DeepSeek's docs date Pro to early August; the expected August 3 date came and went with no release, and the community has since re-coined "early August" as a unit of DeepSeek time. But the staging evidence is now piling on in public:
- An installer script sitting on DeepSeek's own CDN exposed a "Codex DeepSeek Setup v1.0.0" configuration menu — tooling that only makes sense next to a launch.
- The updated pricing documentation states the Responses API is adding V4-Pro support.
- DeepSeek is openly recruiting agent-tool developers for its Harness beta, and its docs quietly dropped the third-party agent products it used to name-check — read widely as clearing the shelf for a first-party harness.
- The August 4 overload waves remain best explained as capacity work.
One number to handle with precision: a leaked "funding memo" circulated on August 6 claiming V4-Pro's coding sits 0.3% behind Claude's flagship. The memo itself was debunked — but DeepSeek's Harness lead clarified where the number actually came from: **Table 6 of the V4-Pro *preview* paper, a specific SWE-bench comparison against Opus 4.6 (80.8 vs 80.6). So the figure is real, the context was stolen: preview-vs-4.6, repackaged as stable-vs-flagship. File it as a measure of the preview, not a preview of the release. Meanwhile the waiting itself has become infrastructure: the community stood up a V4-Pro launch-status site that polls the Responses API around the clock, and the official platform logged another partial outage on August 7 — the fourth week of a serving line running at its limits. And the tracker's longest-running open question is now closed in the worst way: the [officially-previewed peak/off-peak pricing](#july-25-the-aliases-died-on-schedule-the-gray-test-caught-fire) is cancelled** — replaced by a documented notice that prices are being adjusted upward across the board. This page updates when Pro lands.
Signals since July 18 (updated July 20)
Two developments moved the picture, both from the reported-and-community pile:
A bundled coding harness. A DeepSeek harness lead is reported saying the stable V4 will ship *together with* a first-party coding harness — DeepSeek's own agent tooling, launched as a pair. If accurate, that reframes the release: not just a model-card update but DeepSeek's entry into the coding-agent race that Claude Code and Qoder currently define. Community speculation has centered on July 20 as "the last day of mid-July."
The "rebadged Fable" rumor is losing its testers. The gray-test build spawned a theory that stable V4 secretly routes to Claude Fable 5. Community verification has been unusually rigorous and points the other way: side-by-side SVG generation shows materially different output styles; dirty-token probes indicate a tokenizer different from both V3 and the V4 preview; and the visible chain-of-thought summarizes rather than plans in detail. Same-session answers do vary in style, which keeps a *dynamic routing between DeepSeek's own builds* hypothesis alive — but the "it's just Fable" claim has not survived contact with testing. Standard caveat: all of this is community forensics on a gray build, not documentation.
July 23: the release slipped — the deadline didn't
Mid-July came and went. DeepSeek's June 30 signal pointed at a mid-July stable release; a week past that window there is still no announcement, and the community mood has shifted from countdown to "is this slipping again." The most consistent explanation circulating — echoed by the harness-lead report above — is that the release is now gated on the first-party coding harness shipping alongside it: a bundled launch is a bigger bet than a model-card update, and bigger bets slip.
Two things worth separating from the noise. A V4-Pro gray-rollout rumor flared when some users saw expert-mode output styles change; others saw no change at all, so if a gray test exists, its scope is small — file under unverified. And the practical item stands regardless of any of this: the alias retirement does not wait for the stable release. deepseek-chat and deepseek-reasoner stop responding July 24, 15:59 UTC, release or no release. The deadline ships on time even when the launch does not.
July 25: the aliases died on schedule; the gray test caught fire
Two updates since the slip was called.
The deadline executed. deepseek-chat and deepseek-reasoner went dark on July 24 as documented — the one piece of this release that shipped exactly on time. Anything still pointed at them is now failing; migration is a rename.
The gray test became the hottest thing in the community. The week's top thread (hundreds of upvotes) documents a new chain-of-thought format opening with "I am" — matching what a GA build is expected to look like — surfacing in the web app's expert mode, with testers reporting a sharp jump in generation quality over the V4 preview. The folklore has evolved to trigger recipes: off-peak, clean-context single prompts reportedly flip sessions into the new build with high probability. Meanwhile the delay explanation has firmed into community consensus: DeepSeek's harness team reportedly only began hiring in late June, so a bundled model-plus-harness launch waits on the slower half. All of it remains forensics on an unannounced build — but the direction is consistent: the model looks ready; the tooling is the gate.
*July 30:* DeepSeek itself previewed a peak/off-peak API pricing scheme — the exact structure this page has carried as "reported" since day one just moved to officially teased, and the community reads the timing as release choreography. If it ships as previewed, the overseas time-zone math below becomes the practical takeaway.
The overseas angle nobody is pricing in
If the peak-pricing report holds, it quietly favors developers outside China. The reported peak windows — 9:00–12:00 and 14:00–18:00 Beijing time — translate to 01:00–04:00 and 06:00–10:00 UTC, which is nighttime in the US and early morning in Europe. A team in San Francisco running its workload at 10am Pacific lands at 17:00–18:00 UTC (1–2am Beijing): deep off-peak, today's prices. Chinese domestic traffic, concentrated in Beijing business hours, would absorb the 2× windows.
In other words: the reported scheme is congestion pricing for DeepSeek's home market, and most overseas workloads would never touch it. Worth knowing before anyone panics about "DeepSeek doubling prices" — that is not what the report says.
What to do this week (regardless of the release)
The actionable item is the confirmed one. If any of your code, tools or agent configs still reference deepseek-chat or deepseek-reasoner, they stop working on July 24. The migration is a rename:
deepseek-chat→ `deepseek-v4-flash` for routine and high-volume work, or `deepseek-v4-pro` where quality matters.deepseek-reasoner→ `deepseek-v4-pro` in thinking mode.
Both V4 models are OpenAI-compatible, so the change really is one string. Our migration guide covers the mechanics, and the DeepSeek pricing explainer has the current numbers in detail.
What changes at the stable release (and what won't)
Judging by every prior DeepSeek release cycle, a stable label mostly formalizes what the preview already does; weights are already MIT-licensed and the API already serves production traffic at scale. The things to actually watch on release day:
- Whether the time-of-day pricing ships as reported, and its exact windows and multipliers.
- Any bump to rate limits or context/output caps that preview status was holding back.
- Fresh official benchmarks — the preview's coding numbers (LiveCodeBench top score for Pro) were already strong; a stable build that moves them would reshuffle the value rankings.
This page gets updated when the release is official.
Run V4 from anywhere, either way
Both V4 models are live on Turiloop — deepseek-v4-pro and deepseek-v4-flash behind one OpenAI-compatible key, pay-as-you-go with an international card, no Chinese phone number. Point your SDK at api.turiloop.com/v1 and the July 24 alias retirement never touches you; if the stable build changes anything worth switching for, it will be a model-id change, not a migration. The pricing calculator has both models against every rival tier.
FAQ
Is DeepSeek V4 officially released? Half of it: V4-Flash's stable build (0731) shipped July 31, 2026 — retrained, sharply stronger on agent benchmarks, weights open the same day (details). V4-Pro is dated early August; the community's expected August 3 date passed without a launch, and the August 4 official-API overload and maintenance pattern is now read as GA staging.
What happened on July 24, 2026? The legacy deepseek-chat and deepseek-reasoner model aliases were retired at 15:59 UTC and no longer respond. If your code still references them it is failing now — migrate to deepseek-v4-flash or deepseek-v4-pro.
Is DeepSeek doubling its API prices? No — the report says peak-hour pricing: 2× only during Beijing business windows (reported 9:00–12:00 and 14:00–18:00 CST), with current prices outside them. For most overseas workloads those windows fall at night. And it remains a report, not an announcement.
What are DeepSeek V4's specs? Confirmed since April: V4-Pro is a 1.6T-parameter MoE (49B active), V4-Flash is 284B (13B active); both carry 1M context, up to 384K output, thinking and non-thinking modes, MIT-licensed weights.
Should I wait for the stable release to adopt V4? No. The preview already serves production traffic, the weights and prices are public, and the alias deadline forces the naming migration by July 24 anyway. Adopt now; the stable label changes little for API users.