In This Article
Key Takeaways
- On July 31, 2026, DeepSeek's change log announced that “the official release of the DeepSeek-V4-Flash API is now in public beta.” The same entry says the official release of DeepSeek-V4-Pro “will follow soon.”
- There is no DeepSeek V4 stable or general-availability release as of this writing. Both Hugging Face model cards still describe V4 as “a preview version.”
- The July 24, 2026 event was unrelated: the legacy API aliases
deepseek-chatanddeepseek-reasonerwere retired, on a deprecation schedule announced three months earlier. - Published API prices per million tokens: V4-Flash at $0.14 input (cache miss) / $0.28 output; V4-Pro at $0.435 / $0.87. A 2x peak-hour pricing policy is announced but not yet in effect.
DeepSeek's V4 generation has now been in preview for more than three months, and it is still not finished. On July 31, 2026, the company's API change log recorded a single, carefully bounded step forward: “The official release of the DeepSeek-V4-Flash API is now in public beta.” The same entry adds that “the official release of DeepSeek-V4-Pro will follow soon,” and states plainly that “this update only upgrades the DeepSeek-V4-Flash API. The DeepSeek-V4-Pro API and the APP/WEB models are unchanged.” TechNode reported the same the same day, noting that the update “applies only to the V4-Flash API.”
We are stating that carefully because a number of secondary sites have been circulating a different claim — that V4 reached general availability on July 20. DeepSeek's own change log, pricing page, and model cards do not support it, so we do not report it.
What DeepSeek shipped on July 31
The new build is designated V4-Flash-0731. Per the change log, it “keeps the same model architecture and size as DeepSeek-V4-Flash-Preview, and was only re-post-trained” — so this is a training-and-tuning release, not a new architecture. TechNode's account matches: same structure and size as the preview, retrained. Two interface changes ship with it: DeepSeek says the official V4-Flash “natively supports the Responses API format and is specifically adapted for Codex.”
Underneath, the V4 series remains what it was at preview. DeepSeek-V4-Pro is a mixture-of-experts model with 1.6 trillion total parameters and 49 billion active; DeepSeek-V4-Flash has 284 billion total and 13 billion active. Both carry a one-million-token context window, and both model repositories are licensed under MIT. Those figures appear on the V4-Pro and V4-Flash model cards and were reported independently by TechCrunch at launch.
The timeline, corrected
Three dates, in order, all from primary or independently corroborated sources:
April 24, 2026. DeepSeek published the V4 series as a preview — V4-Pro and V4-Flash, open weights under MIT. TechCrunch described them that day as “two preview versions.” Bloomberg covered the release under the headline “DeepSeek Unveils Newest Flagship AI Model a Year after Upending Silicon Valley.” The same change log entry set a deprecation clock: the legacy names deepseek-chat and deepseek-reasoner “will be discontinued in three months (2026-07-24).”
June 30, 2026. TechNode reported, citing an announcement from the DeepSeek team, that “the official release of DeepSeek V4 is scheduled for mid-July,” alongside a new peak-hour API pricing mechanism.
July 24, 2026. The two legacy aliases were retired on schedule. This was a deprecation deadline, not a release. Some write-ups give a precise cutoff time; we could not confirm one from any primary source, so we omit it.
Mid-July came and went
The publicly stated target was an official V4 release in mid-July. What arrived, on July 31, was one of the two models, labeled public beta, with the larger model's date unannounced. That is a slipped and staggered rollout, and it is worth naming as such rather than filing it as a launch.
The benchmark numbers, and who reported them
Two scores for V4-Flash-0731 appear in both DeepSeek's change log and TechNode's independent write-up: Terminal-Bench 2.1 at 82.7 and DeepSWE at 54.4. Those are the two we would treat as best-sourced.
DeepSeek's change log also publishes a wider set that we found only there, and which we therefore present as the company's own reported figures rather than as verified results: NL2Repo 54.2, Cybergym 76.7, Toolathlon (verified) 70.3, DSBench-FullStack 68.7, DSBench-Hard 59.6, Agent Last Exam 25.2, and Automation Bench (Public) 25.1. DeepSeek's accompanying claim is that the build has “significantly enhanced agent capabilities, with benchmark results far exceeding V4-Pro-Preview” — that is, the smaller and cheaper model now outscores the company's own larger preview flagship on agent work. It is a vendor claim, and a striking one; treat it as a reason to test, not as a settled result.
One third-party reading is available: Artificial Analysis scores DeepSeek V4 Flash 0731 at 50 on its Intelligence Index.
Pricing, and a surge scheme that is not live
All figures below are from DeepSeek's published Models & Pricing page, in USD per one million tokens, as of July 31, 2026.
DeepSeek V4 API pricing, per 1M tokens
| Model | Input (cache hit) | Input (cache miss) | Output |
|---|---|---|---|
| deepseek-v4-flash | $0.0028 | $0.14 | $0.28 |
| deepseek-v4-pro | $0.003625 | $0.435 | $0.87 |
Both models list a 1M-token context window and 384K maximum output. Two details are worth flagging. First, the pricing page attaches no “preview” or “beta” label to either model, which sits oddly next to a change log that calls Flash a public beta and model cards that still say preview. Second, the peak-hour policy is announced but not active: the page describes prices at 2x during 9:00–12:00 and 14:00–18:00 Beijing time (UTC+8), “applicable to all billing items,” then says “the effective date will be subject to the official announcement.” Anyone telling you surge pricing is live today is ahead of the record.
On price history, we report only what DeepSeek publishes: $0.435 / $0.87 for V4-Pro as current standing rates, no promotional label attached. Secondary write-ups describe a large introductory discount off higher April list prices and disagree on when it became permanent; we could not confirm that from DeepSeek's own materials.
Why it matters
The section above is reported fact. What follows is our analysis.
For practitioners, the useful signal here is not the benchmark table — it is the label. “Public beta” and “preview” are not decoration; they are the vendor telling you the contract is not yet firm. Behavior can change under a re-post-training, as it just did for Flash, and a model whose name stays the same while its weights are retrained will quietly move your evaluation scores. If you are running V4 in anything that matters, pin your prompts to a dated build where you can, keep a regression set, and re-run it after every change-log entry.
The second point is about routing. If DeepSeek's own agent numbers hold up on your workload, the cheaper Flash tier is the one to test first, and the 3x price gap to Pro stops being an obvious upgrade path. That is the same discipline we argue for across model families in choosing a model tier in 2026: start at the bottom, escalate only on failure. The cache-hit pricing sharpens it further — at a 50x spread between cache hit and miss, how you structure a prompt can matter more to your bill than which tier you picked, a dynamic we cover in context economics.
Third, the announced-but-dormant peak-hour policy is a cost variable to plan for before it activates. Time-of-day pricing at 2x, applied to all billing items, changes the economics of batch jobs and scheduled agent runs far more than it changes interactive use. If your pipeline fires at a fixed hour, it is worth knowing now whether that hour falls inside the published peak window — and worth modeling the delta on our LLM price calculator before the effective date lands.
Finally, both V4 repositories are MIT-licensed, so the API's release status is not your only path to the models. For teams weighing hosted inference against running weights themselves, our guide to open-weight versus frontier API lays out where each choice wins.
Price the tiers against your own token volume
Compare DeepSeek V4-Flash and V4-Pro against the other model families on your real input and output mix — including cache-hit ratios — before you commit a pipeline to one of them.
Open the price calculatorSources: DeepSeek API Change Log; DeepSeek Models & Pricing; DeepSeek V4 release notes (April 24, 2026); TechNode, July 31, 2026; TechNode, June 30, 2026; TechCrunch, April 24, 2026; DeepSeek-V4-Pro model card; DeepSeek-V4-Flash model card; Artificial Analysis; Bloomberg, April 24, 2026 (headline). Analysis and framing by Precision AI Academy.
Common questions
Is DeepSeek V4 out of preview? Not fully. The V4-Flash API is in public beta as of July 31, 2026. V4-Pro's API is unchanged, and both Hugging Face model cards still describe the series as a preview version.
Do I need to change my API calls? Only if you were still using the old aliases. deepseek-chat and deepseek-reasoner were retired on July 24, 2026, per a schedule published on April 24. The current identifiers are deepseek-v4-flash and deepseek-v4-pro.
Did the model get bigger or change architecture? No. DeepSeek states that V4-Flash-0731 keeps the same architecture and size as the preview build and was only re-post-trained.
When does V4-Pro get its official release? Unannounced. DeepSeek says only that it “will follow soon.” Separately, the pricing page indicates Responses API support for deepseek-v4-pro is anticipated in early August 2026 — a forward-looking statement in DeepSeek's docs, not a shipped feature.