DeepSeek V4 Release (2026): Specs, Price & Verdict

Back in July everyone was refreshing the DeepSeek changelog daily, arguing about whether V4 was “confirmed” or “rumor.” That debate is over. As of September 2026 the DeepSeek V4 release is not a leak, a teaser, or a Reddit thread — it is a shipped, generally available model family with published prices, weights on Hugging Face, and a newer V4.1-Flash already in production. So this post does what the situation now demands: it lays out exactly what DeepSeek V4 is today, what it costs, how it benchmarks against GPT-5.4, Gemini 3.1 Pro and Claude Opus, and the one pricing detail that will quietly change your bill.

How we assess: this article is based on DeepSeek’s official API changelog, published pricing pages, and verified developer reports — not hands-on lab testing. Last updated September 2026.

What is DeepSeek V4?

DeepSeek V4 is the follow-up to V3.2 from the Chinese lab that upended AI pricing in 2025. It is an open-weight, MIT-licensed model family with a 1-million-token context window, released in two core sizes: V4-Pro, the flagship, and a smaller, faster Flash tier. The preview launched April 24, 2026 with weights posted to Hugging Face, and the stable V4-Pro reached general availability across DeepSeek’s app, web, and API on August 13, 2026. The architecture is Mixture-of-Experts: V4-Pro carries about 1.6 trillion total parameters but activates roughly 49 billion per token, which is how it stays cheap to serve despite its size.

The reason this launch matters is not that DeepSeek built the single best model on earth — it did not. It is that DeepSeek got within striking distance of the frontier labs on coding and math while charging a small fraction of their price, and then handed out the weights. That combination is what keeps forcing the rest of the market to cut prices.

Who this model is for

V4-Pro is aimed at developers and teams who run high-volume coding, agentic, and reasoning workloads and care intensely about cost per token. If you are burning through millions of tokens a day on code generation, refactoring, or automated agents, V4-Pro’s price-to-performance is hard to ignore. The Flash line (now V4.1-Flash) targets high-throughput, latency-sensitive jobs: classification, extraction, chat, and multimodal tasks where you want frontier-ish quality at commodity pricing. If you need the absolute best reasoning for a hard, low-volume problem, you will still reach for Claude Opus or Gemini — but for everything in the “good enough, at scale, cheap” bucket, this is the model to beat. For a direct money comparison, see our breakdown of DeepSeek V4 vs GPT-5.6 on price.

DeepSeek V4 specs and pricing at a glance

Here is the current state of the family as of September 2026. Prices are per million tokens; the Flash line uses peak/off-peak billing, so both numbers are shown.

Is DeepSeek V4 officially released?

Yes, unambiguously. The timeline no longer has any “expected” dates in it. The preview shipped April 24, 2026. V4-Pro reached general availability on August 13, 2026 with three selectable thinking-effort levels — low, high, and max — and meaningfully upgraded agent behavior. On August 21 DeepSeek added an experimental vision model, V4-Flash-Vision-Exp, whose multimodal agent performance the lab described as approaching Opus-4.8 level. Then on September 10, V4.1-Flash arrived as the deepseek-flash endpoint with native image understanding, higher throughput, and lower prices, and the older Flash names were pointed at it. If you were waiting for “confirmed,” the wait ended weeks ago.

How good is it? The benchmarks, honestly

On code and math, DeepSeek V4-Pro is genuinely frontier-adjacent. On the top-effort configuration DeepSeek reported for the flagship, SWE-bench Verified came in at 80.6%, a whisker behind Claude Opus 4.6’s 80.8%. LiveCodeBench actually went the other way, with V4-Pro at 93.5 against Opus at 88.8, and its Codeforces rating of 3206 edged out GPT-5.4’s 3168 and Gemini 3.1 Pro’s 3052. On competition math it is elite: HMMT 2026 at 95.2% (Opus 4.6: 96.2%) and a clean 120/120 on Putnam 2025.

Now the honest counterweight. On broad general knowledge and hard science reasoning, it trails the very top. MMLU-Pro landed at 87.5% versus Gemini 3.1 Pro’s 91.0%, and GPQA Diamond at 90.1% versus Gemini’s 94.3%. My read: for coding, agents, and math, V4-Pro competes with anything and undercuts all of it on price. For the hardest reasoning, research, and knowledge tasks, Gemini and the top Claude models still have a real edge. Pick the model to the job, not the headline. If you want the full frontier picture, our GPT-5.4 vs Claude Opus head-to-head is a useful companion read.

How much does the DeepSeek V4 API cost?

This is where DeepSeek keeps winning. V4-Pro is priced at roughly $0.435 per million input tokens and $0.87 per million output, with cache-hit input around $0.0036. V4.1-Flash, the current default for high-throughput work, runs about $0.30/$1.20 per million input/output at peak and half that off-peak. Put that against closed frontier output pricing that runs into the tens of dollars per million tokens and the gap is not small — it is an order of magnitude or more. For a lot of production workloads, switching a bulk task from a premium model to V4 is the single biggest line-item cut you can make this quarter. [ADD YOUR EXPERIENCE: if you have moved a real workload onto V4, drop the before/after monthly cost here — readers trust concrete numbers far more than benchmark charts.]

How do you access DeepSeek V4?

There are four realistic paths, depending on what you are doing. The fastest is the consumer chat at chat.deepseek.com, which is free for casual use and runs the current models — fine for trying it before you wire anything up. For production, you register on the DeepSeek API platform, generate a key, and call the OpenAI-compatible endpoint with the model name deepseek-v4-pro or deepseek-flash; the compatibility layer means most existing OpenAI SDK code works with a base-URL swap. If you want the weights, they are on Hugging Face under MIT for self-hosting or fine-tuning. And if data residency rules keep you off the first-party API, several Western inference providers now host the open weights, so you get V4 quality without sending traffic to DeepSeek’s own servers. Match the path to your governance needs, not just your budget.

What no one else tells you about the DeepSeek V4 release

Three things get glossed over in the hype coverage, and they matter more than any benchmark.

The peak/off-peak clock is the real pricing story. Those headline-cheap Flash numbers are the peak rate; the off-peak rate is half. But “off-peak” is defined in UTC, and peak hours run roughly 01:00–04:00 and 06:00–10:00 UTC on weekdays. For a US team, a chunk of your normal workday sits inside DeepSeek’s peak window, so the price you actually pay is not always the lowest one advertised. If your workload is batchable, scheduling it into off-peak UTC hours can literally halve your bill. Almost nobody prices this in.

“Open weight” does not mean “you can run it.” Yes, V4-Pro’s 1.6-trillion-parameter weights are on Hugging Face under MIT. No, you are not running that on a workstation, or even a single 8-GPU box, at full precision. Self-hosting the flagship is a serious infrastructure project. For almost everyone, “open weight” means freedom from lock-in and the option of a third-party host — not a laptop deployment. The Flash models are far more realistic to self-host, and that is where the open license actually pays off for smaller teams.

Near-parity on a benchmark is not parity in your product. V4-Pro scoring 80.6% to Opus’s 80.8% on SWE-bench tells you they are close on that test. It does not tell you they feel identical inside your agent, on your codebase, with your prompts. Benchmarks compress a thousand judgment calls into one number. Run your own eval on your own tasks before you migrate anything that matters. This is not a knock on DeepSeek — it is true of every model launch, and the labs bank on you forgetting it.

DeepSeek V4 vs the alternatives

Where does V4 actually fit against the other options you would realistically consider?

If you are building a coding agent, the practical move is to pair V4-Pro’s cheap tokens with a strong editor — our Cursor review covers the setup most developers land on.

Who should adopt V4 — and who should wait

Adopt it now if you: run high-volume coding or agent workloads, are cost-sensitive, want to avoid vendor lock-in, or need a 1M-token context on a budget. Best for indie developers, startups watching burn, and platform teams processing tokens by the billion.

Wait or look elsewhere if you: need the single best score on hard reasoning and research, have strict data-governance rules about Chinese-hosted APIs (self-host the open weights or use a trusted third-party host instead), or want a mature enterprise support contract with guarantees. Best avoided as a drop-in for mission-critical reasoning where a few points of accuracy justify paying 20× more.

The honest limits

Let me be blunt about the caveats. Vendor-reported benchmarks always flatter the vendor, so treat the numbers as directional. The peak/off-peak pricing adds real complexity to cost forecasting. Hosting policy and data residency are genuine concerns for regulated industries using the first-party API. Caching is described as “best effort,” so do not build a cost model that assumes guaranteed cache hits. And the family is moving fast — V4-Pro is already being folded toward the Flash line, so pin your model names and watch the changelog, because “current” in this space has a shelf life measured in weeks. None of this erases the value. It just means you should run your own eval and model your own bill before committing.

Frequently asked questions

Is DeepSeek V4 officially released?

Yes. The preview launched April 24, 2026, V4-Pro reached general availability on August 13, 2026, and V4.1-Flash shipped September 10, 2026. It is a fully released, generally available model family, not a preview or rumor.

How much does the DeepSeek V4 API cost?

V4-Pro is about $0.435 per million input tokens and $0.87 per million output. V4.1-Flash is roughly $0.30/$1.20 per million input/output at peak and half that off-peak. Cache-hit input is a fraction of a cent per million.

What is DeepSeek V4’s context window?

One million tokens across the V4 family, including V4-Pro and V4.1-Flash. That is enough for very large codebases, long documents, or extended agent histories in a single call.

Is DeepSeek V4 open source?

The weights are open and MIT-licensed, published on Hugging Face, so it is open-weight. That allows self-hosting and commercial use, though the training data and full pipeline are not released, so “open weight” is more precise than “fully open source.”

What is the difference between V4-Pro, V4-Flash, and V4.1-Flash?

V4-Pro is the 1.6-trillion-parameter flagship for the hardest coding and reasoning. V4-Flash was the smaller, cheaper preview tier. V4.1-Flash, released September 2026, replaces it with native multimodal vision, higher throughput, and lower prices, served at the deepseek-flash endpoint.

What are DeepSeek’s peak and off-peak hours?

Peak hours run roughly 01:00–04:00 and 06:00–10:00 UTC on weekdays, with off-peak billing set at half the peak rate. Because the window is defined in UTC, part of a US workday falls inside peak pricing.

How does DeepSeek V4 compare to GPT-5.4 and Claude Opus?

On coding and math it is competitive: V4-Pro scored 80.6% on SWE-bench Verified versus Opus 4.6’s 80.8%, and led on LiveCodeBench and Codeforces. On broad knowledge (MMLU-Pro, GPQA Diamond) it trails Gemini 3.1 Pro. It wins decisively on price.

Can I run DeepSeek V4 on my own hardware?

The Flash models are realistic to self-host on capable GPU setups. V4-Pro’s 1.6-trillion-parameter weights, while open, need data-center-class hardware to run at full precision, so most teams use the API or a third-party host rather than running the flagship locally.

Is DeepSeek V4 safe for business use?

Technically it is production-ready and widely used. The real questions are governance and data residency: regulated organizations often prefer to self-host the open weights or use a trusted third-party host rather than send data to the first-party API. Run it past your compliance team before adopting.

Should I use V4.1-Flash instead of V4-Pro?

For high-throughput, latency-sensitive, or multimodal tasks, yes — V4.1-Flash is cheaper and faster and now handles the legacy Flash traffic. Keep V4-Pro for the hardest coding, agentic, and reasoning jobs where you want the flagship’s extra capability.