DeepSeek V4: Price, Specs, and Why It Is Still a Preview (August 2026)

Last updated: August 2026

DeepSeek V4 is a two-model family — V4-Pro at 1.6 trillion total parameters and V4-Flash at 284 billion — released as an open-weight preview on April 23, 2026 under the MIT License. As of August 4, 2026, V4-Pro remains a preview, V4-Flash's official API is in public beta, and V4-Pro output tokens cost $0.87 per million: 75% below the launch price this article originally reported.

This page now works as a status ledger. The release moved fast enough that most coverage of it — including the first version of this article — went stale within weeks. Every claim below carries a date and a source, so you can see exactly what changed and when.

What Changed Since April: The Ledger

Five events have materially changed what DeepSeek V4 is since launch day. If you evaluated V4 in the spring and moved on, this table is the catch-up.

All rows verified against DeepSeek's official documentation and independent pricing trackers on August 4, 2026.

Is DeepSeek V4 Actually Released? The Honest Status

No — not fully, and this is where most coverage gets it wrong. A wave of roundups published in late July declared that DeepSeek V4 reached general availability on July 20, 2026. DeepSeek's own documentation does not support that claim.

Here is what the primary sources show. The July 31 entry in DeepSeek's API changelog announced an official release for V4-Flash only, explicitly labeled a public beta, and closed by promising that an official V4-Pro release would follow — in the changelog's words, "The official release of DeepSeek-V4-Pro will follow soon." A company does not promise a forthcoming official release for a model that already went GA eleven days earlier. Meanwhile, the V4-Pro repository on Hugging Face still opens with the sentence "We present a preview version of DeepSeek-V4 series."

The July 20 GA claims trace back to community reports of grayscale testing — small-batch trials of candidate builds on Chinese platforms — plus leaked internal build identifiers (deepseek-v4-pro-202606 among them) that circulated in early July. Grayscale testing is real evidence that a GA build exists. It is not a release. Chinese tech media have since reported a possible official window of August 10 to 20, tied to a closed test of DeepSeek's in-house Harness coding tool; treat that as reported, not confirmed, because DeepSeek has announced no date.

The practical read: the V4-Pro you can call today is the April preview checkpoint at permanent post-cut pricing. The V4-Flash you call today is a newer build than the one on Hugging Face. Both facts matter, and neither appears in a headline that just says "released."

V4-Pro and V4-Flash: Verified Specifications

The core specifications have held stable since launch and are drawn from DeepSeek's model cards. Both models are Mixture-of-Experts (MoE) architectures: a large total parameter count with only a small fraction active on any given token.

Specifications from DeepSeek's official model cards, checked August 4, 2026. At launch, V4-Pro overtook Kimi K2.6 (1.1 trillion) as the largest open-weight model released, per TechCrunch's release-day reporting.

Why V4 Is Architecturally Interesting

The consequential part of this release is not the parameter count. It is how much compute and memory DeepSeek removed from long-context inference — which is the engineering story underneath the pricing story, and the reason the May price cut is structural rather than promotional. In DeepSeek's published figures, at a 1 million token context V4-Pro uses roughly 27% of the single-token inference compute and 10% of the KV cache memory of its V3.2 predecessor; V4-Flash pushes that to roughly 10% and 7%.

Four choices drive the gains. A hybrid attention design pairs Compressed Sparse Attention with Heavily Compressed Attention, so long contexts are processed without comparing every token to every other token at full precision. Manifold-Constrained Hyper-Connections stabilize signal propagation at extreme depth. The Muon optimizer — first popularized by Moonshot AI's Kimi series — handles training stability. And quantization-aware training in mixed FP4 and FP8 precision keeps the on-disk footprint in the 160 GB to 865 GB range rather than multiple terabytes.

One supply-chain note has aged well: DeepSeek validated V4's expert-parallel scheme on both Nvidia GPUs and Huawei Ascend NPUs, continuing the pattern we covered when Z.ai trained GLM-5.1 entirely on Huawei silicon. The Chinese open-weight frontier increasingly does not depend on Nvidia hardware, which matters for export-control watchers.

What DeepSeek V4 Costs Now

This is the section that made the refresh necessary. The launch prices this article originally reported — $1.74 input and $3.48 output for V4-Pro — lasted five weeks. On May 31, 2026, DeepSeek made its promotional 75% cut permanent, a move InfoWorld characterized as an escalation of the AI pricing war. The rates below are the current standard rates, verified against DeepSeek's pricing page and independent trackers on August 4, 2026, alongside the closed-frontier list prices in effect the same day.

USD per million tokens, list prices as of August 4, 2026. Gemini rate applies to prompts up to 200K tokens. Competitors offer their own caching and batch discounts under different schemes, so cache-hit columns are shown only where directly comparable. Token counts also differ by tokenizer: Anthropic's current tokenizer produces roughly 30% more tokens for the same text than earlier generations, which narrows any per-token comparison.

The arithmetic on cache-miss output pricing: V4-Pro at $0.87 is roughly one twenty-ninth the price of Claude Opus 5, one thirty-fourth of GPT-5.5, and one fifty-seventh of Claude Fable 5. V4-Flash at $0.28 sits below one one-hundredth of the Fable 5 rate. When this article launched, the comparable gap was one-seventh — the May cut widened it fourfold, which is why quoting an undated AI price is quoting a wrong one.

Two forward-looking caveats. First, thinking-mode reasoning tokens bill at the output rate, so reasoning-heavy calls cost more than the headline implies. Second, DeepSeek has announced a 2x peak-hour surcharge during Beijing business hours but had published no effective date as of August 4. Check the official rate card before committing production budgets; this market has already repriced itself once this year.

Benchmarks: What Is Verified and What Is Not

DeepSeek's technical report describes V4 as trailing the closed frontier by roughly three to six months on the hardest cross-domain reasoning — an unusually honest self-assessment worth anchoring before any specific number. The vendor-reported picture: V4-Pro's 80.6% on SWE-bench Verified is the strongest open-weight score on real-world software engineering, essentially tied with the Claude Opus generation it launched against, and its competitive-programming and structured-math results led the field at release. Artificial Analysis ranked V4-Pro among the top open-weight reasoning models in its independent index.

The caveat has grown teeth since April. DeepSeek's newer agent-benchmark figures — including the July 31 claims that the updated V4-Flash now out-scores V4-Pro's preview on agentic suites — were measured with the company's own evaluation harness, which remains unpublished, on test sets that partially remain internal. No one outside DeepSeek can currently reproduce the full table. Our breakdown of the V4-Flash-0731 release reads those asterisks in detail. Treat vendor benchmarks as claims with a confidence interval, not as measurements, until independent evaluators replicate them.

The Data Sovereignty Question

Using DeepSeek's hosted API means your prompts and outputs traverse Chinese infrastructure. For most personal use, that is not categorically different from trusting OpenAI or Anthropic servers in the United States — you are handing tokens to a foreign operator either way. For regulated industries, IP-sensitive work, or data-residency obligations, it is a harder question, and DeepSeek's history adds context: Anthropic's February 2026 distillation report named it among the Chinese labs accused of extracting Claude's capabilities through proxy accounts. That allegation does not make V4 less technically impressive; it is simply part of the record when deciding where your prompts go.

The July 31 release also handed sovereignty advocates a concrete exhibit: the model served under the deepseek-v4-flash API name changed overnight — same name, new checkpoint, no user action — while Hugging Face continued hosting the April weights. An API model's identity belongs to its operator. The mitigations remain what they were: self-host V4-Flash so the weights on your disk cannot change without you, route through a non-DeepSeek host such as OpenRouter after reading its data-handling terms, or split traffic so only non-sensitive workloads capture the price advantage. The MIT license is what keeps the first option real.

Can You Run V4 at Home?

V4-Pro, no: at roughly 865 GB it is a multi-GPU data-center deployment, an open-weight enterprise model rather than a local one. V4-Flash, yes, with real hardware requirements: mainline llama.cpp gained V4 support in July 2026, and the working floor is about 110 GB of memory, putting the recommended quantized build on 128 GB unified-memory machines or 96 GB GPU rigs. The full memory math, quantization tiers, and buying guidance live in our V4-Flash hardware reality check. If your machine sits under that floor, run a model that fits well instead — our guides to what AI models actually run at home and the best hardware for local AI map the tiers.

Frequently Asked Questions

Is DeepSeek V4 fully released or still a preview?

Partially released. As of August 4, 2026, V4-Pro remains the April preview build; DeepSeek's July 31 changelog promises its official release is forthcoming. V4-Flash received an official API release on July 31, 2026, explicitly labeled a public beta. Widely repeated claims of a July 20 general availability trace to community grayscale-test reports, not to any DeepSeek announcement, and Chinese media reports of an August 10 to 20 window are unconfirmed.

What does the DeepSeek V4 API cost right now?

As of August 4, 2026, V4-Pro costs $0.435 per million input tokens on a cache miss, $0.003625 on a cache hit, and $0.87 per million output tokens. V4-Flash costs $0.14, $0.0028, and $0.28 respectively. These are the permanent rates set on May 31, 2026, when DeepSeek converted its 75% promotional discount into standard pricing. A 2x peak-hour surcharge has been announced but has no effective date yet.

What is the difference between V4-Pro and V4-Flash?

V4-Pro is the flagship: 1.6 trillion total parameters with 49 billion active per token. V4-Flash is the efficient sibling: 284 billion total with 13 billion active. Both share the 1 million token context window, 384K maximum output, three reasoning modes, and the MIT license. Since the July 31 update, DeepSeek's own agent benchmarks show the refreshed Flash ahead of the Pro preview on agentic tasks — vendor-reported, so verify against your workload — which makes Flash the default starting point for most users.

Can I run DeepSeek V4 on my own hardware?

V4-Pro requires data-center hardware and is not a home deployment at any budget an individual should entertain. V4-Flash runs locally as of July 2026 through mainline llama.cpp with a memory floor around 110 GB, which means a 128 GB unified-memory machine or a 96 GB multi-GPU rig. Below that floor, smaller open models such as Gemma 4 and Qwen variants are the practical choices; our hardware reality check covers the full tier map.

What happened to deepseek-chat and deepseek-reasoner?

Both legacy model names were retired on July 24, 2026 at 15:59 UTC, and API requests to them now fail. During the transition period, DeepSeek routed them to V4-Flash in Non-Thinking and Thinking modes. Any integration still pointing at the old strings must move to deepseek-v4-pro or deepseek-v4-flash; the change is a model-string swap, since the API supports the same OpenAI-compatible calling conventions.

Is it safe to send sensitive data to DeepSeek's API?

It depends on your threat model. Prompts to the hosted API traverse Chinese infrastructure, which for personal use is comparable to trusting any foreign operator, but for regulated or IP-sensitive work is a genuine data-residency concern — compounded by the fact that the model behind an API name can change without notice, as happened on July 31. The mitigations are self-hosting V4-Flash so your data never leaves your network, using a non-DeepSeek host with acceptable data-handling terms, or routing only non-sensitive traffic to V4.