DeepSeek V4 Pro Is Finally Official — And the Price Hike Is No Joke

DeepSeek V4 Pro Is Finally Official — And the Price Hike Is No Joke



On Thursday, August 13, 2026, Reuters reported that DeepSeek had formally released the official version of its V4 Pro model. If you work in tech, code with AI tools, or just keep half an eye on the AI arms race between the U.S. and China, this is one of those "wait, already?" moments. DeepSeek V4 Pro had been sitting in preview since April. Now it's live, it's meaningfully better at agent tasks, and — in classic DeepSeek fashion — it arrived with almost no fanfare and one very large catch: prices are about to jump, in some cases by over 1,000%.

This piece breaks down exactly what shipped, what the benchmark numbers actually mean (and don't mean yet), how the new pricing compares to Claude, GPT, and Grok, what developers on Reddit and Hacker News are saying after hands-on testing, and what all of this means if you're a North American business or developer deciding whether DeepSeek belongs in your stack.

Table of Contents

What Just Happened With DeepSeek V4 Pro

Let's start with the plain facts, because the story here is easy to overstate in either direction.

DeepSeek first showed V4 and V4 Pro to the world on April 24, 2026, but that release carried a "preview" label. For nearly four months, developers building on top of it knew the model could change under their feet at any time. On August 12–13, 2026, that changed. DeepSeek quietly updated its official API pricing page to point the existing deepseek-v4-pro endpoint at a new production build called DeepSeek-V4-Pro-0813. Reuters confirmed the formal release the same day, noting the company is trying to regain ground against faster-moving domestic rivals while it expands hiring, computing capacity, and fundraising.

There was no press release. No blog post announcement. No changelog entry marking the swap. AI researcher Simon Willison, who tracks model releases closely, pointed out on Hacker News that the only real signal was a documentation update on DeepSeek's own pricing page, first spotted by developers testing on OpenRouter. That low-key rollout is very on-brand for DeepSeek, but it also means a lot of casual users have no idea this happened yet — which is exactly why it's worth explaining clearly.

Here's the headline: model name and API endpoint are unchanged, so existing integrations keep working without code changes. What changed underneath is the training. DeepSeek says the 0813 build focuses heavily on agent capabilities — the kind of work where an AI system uses tools, runs code, browses files, and completes multi-step tasks with minimal hand-holding.

Why This Matters Right Now

Three separate storylines are colliding at once, and that's what makes this release worth paying attention to instead of scrolling past.

First, the "little sibling problem." DeepSeek's cheaper V4 Flash model reached general availability on July 31, 2026, and immediately started beating the April preview of V4 Pro on several independent agent benchmarks. That's an awkward look for any AI company — your budget model outperforming your flagship isn't supposed to happen. The V4 Pro 0813 release reads like DeepSeek's direct answer to that problem, and the benchmark jump backs that reading up.

Second, the money. DeepSeek raised roughly $7.4 billion in its first outside funding round in June 2026, a notable reversal for a company that had spent its whole existence avoiding outside capital. Reuters reported in July that DeepSeek was already back at the table discussing a new round that could value the company near $74 billion. Training and serving frontier models at scale is extraordinarily expensive, and this release lands right in the middle of that fundraising push.

Third, the pricing shock. DeepSeek built its entire reputation on being the cheap, disruptive alternative to Western AI labs. This release comes bundled with an announcement that API prices are about to rise — in some cases by more than 10x. For a company whose whole pitch has been "frontier performance at a fraction of the cost," that's a meaningful shift, and it's the part of this story with the most practical impact on anyone actually building products on top of DeepSeek.

Inside V4 Pro 0813: What Actually Changed

DeepSeek V4 Pro is a Mixture-of-Experts (MoE) model — 1.6 trillion parameters total, with about 49 billion active per token. That architecture hasn't changed between the preview and 0813. What did change:

  • Reasoning effort controls: Developers can now pick between "low," "high," and "max" reasoning effort, with DeepSeek recommending the middle setting for everyday agent work.
  • Native OpenAI Responses API support: V4 Pro now plugs directly into the Responses API format, with built-in Codex integration — a meaningful convenience if you're already building on OpenAI-style tooling.
  • Anthropic-compatible endpoint: Tools built for Claude, including Claude Code-style agents, can point at DeepSeek's Anthropic-format endpoint with minimal rework.
  • "Expert Mode" in the consumer app: The V4 Pro model is now accessible inside DeepSeek's chat app and web interface under a dedicated Expert Mode toggle.
  • Speculative decoding (DSpark): Under-the-hood inference improvements that speed up generation without changing output quality.
  • Unchanged specs: 1-million-token context window, up to 384,000 tokens of output, and support for both "thinking" and "non-thinking" response modes.

According to DeepSeek's own comparison chart, circulated first through the company's WeChat channel and then picked up by developers on X and Hacker News, the model can also be run through OpenRouter, which listed the 0813 build on August 12, 2026.

The Benchmark Numbers — And the Big Asterisk

This is the part of the story that deserves the most scrutiny, so let's slow down here.

DeepSeek's own benchmark table shows large jumps from the April preview to the 0813 release:

  • Terminal-Bench 2.1: 72.1 → 87.9
  • DeepSWE: 12.8 → 62.7
  • CyberGym: 52.7 → 83.3
  • NL2Repo: 61.5
  • Toolathlon-Verified: 74.1
  • DSBench-FullStack: 71.1

Framed a different way: on Terminal-Bench 2.1, DeepSeek's own numbers put V4 Pro 0813 just 0.1 points behind Anthropic's Fable 5 (88.0) and roughly 2.9 points ahead of Claude Opus 4.8. That's a striking claim for a model priced dramatically lower than either. One analysis by explainx.ai calculated the new build lands within striking distance of Fable 5 at somewhere around a fraction of the per-token output cost.

Here's the asterisk, and it's an important one: every one of these figures is vendor-reported. As of publication, no independent lab had verified them. Third-party benchmark tracker benchable.ai's page for the model reportedly showed an empty results section — nothing had been reproduced outside DeepSeek yet. DeepSeek's own official changelog didn't even have an entry for the 0813 release at launch. And the open weights that would let outside researchers actually test the model themselves had not been published, even though DeepSeek has a track record of eventually releasing open weights for its models.

None of that means the numbers are wrong. DeepSeek has generally been credible with past releases. But "vendor reports big benchmark gains against its own previous model" and "independent labs confirm those gains" are two very different levels of evidence, and right now we're only at the first one.

A Word on How to Read Vendor Benchmarks

One nuance worth flagging: some social media posts described the DeepSWE improvement as a "+389% jump." That's a relative-percentage framing that makes the gain sound larger than it is. The more honest way to describe a move from 12.8 to 62.7 is a 49.9 percentage-point increase — still a genuinely big improvement, just not the eye-popping figure some headlines implied. When you see AI benchmark numbers thrown around online, it's worth checking whether you're looking at percentage points or a relative percentage change. They tell very different stories.

The Price Hike Everyone Is Talking About

Here's where this story gets genuinely newsworthy for anyone actually paying a DeepSeek invoice.

At launch, pricing stayed exactly where the preview left it: $0.435 per million input tokens on a cache miss, $0.003625 per million on a cache hit, and $0.87 per million output tokens. That's still remarkably cheap by frontier-model standards.

But DeepSeek announced, alongside the launch, that a new pricing structure takes effect at 00:00 Beijing time — reported by various outlets as landing on either August 16 or 17, 2026 depending on how the time zone converts locally. The company is introducing peak and off-peak billing for the first time:

  • Peak hours: 09:00–12:00 and 14:00–18:00 Beijing time
  • Off-peak hours: everything else, priced at roughly half the peak rate
  • Overall increase range: reported between 50% and 1,100%, depending on the specific model, token type, and time of day
  • Output tokens (V4 Pro): rising from a flat $0.87 per million to as high as $3.96 per million during peak hours, according to Nikkei Asia's reporting citing Reuters
  • Cache-hit input: the single largest jump — up to 12 times the previous rate during peak hours, and roughly 6 times during off-peak

Even at the new peak rate, DeepSeek's pricing remains well below several Western competitors — Nikkei Asia's coverage notes Anthropic's Fable 5 charges $50 per million output tokens, more than 12 times DeepSeek's new peak rate. But the direction of travel matters. DeepSeek built its entire market position on undercutting everyone else by a wide margin. This is the company acknowledging, in effect, that its old pricing wasn't sustainable at scale.

It's also worth remembering DeepSeek had already reversed course once before: a 75% promotional discount that launched as a temporary offer through May 5, 2026 was later made permanent pricing, only for this new hike to arrive a few months later. If you're budgeting around DeepSeek's API for a production application, this is not a "set it and forget it" situation — pricing here has moved twice in under four months.

How V4 Pro Stacks Up Against the Competition

[Comparison Table]

Model Input Price (per 1M tokens) Output Price (per 1M tokens) Context Window Best For
DeepSeek V4 Pro 0813 $0.435 (pre-hike) $0.87 (pre-hike); up to $3.96 peak post-hike 1M tokens Budget-conscious agentic coding, high-volume workflows
DeepSeek V4 Flash $0.14 $0.28 1M tokens Lightweight agent tasks, cost-first pipelines
Claude Opus 4.8 Higher tier pricing Higher tier pricing Large context Enterprise reasoning, coding accuracy, safety-sensitive use
Claude Fable 5 Premium tier $50 Large context Top-tier agentic and coding performance
Grok 4.6 (SpaceXAI) $2 $6 Large context Long-running agents, coding, visual/interactive tasks

Pricing reflects publicly reported figures as of August 13, 2026. Always confirm current rates on each provider's official pricing page before committing production workloads, since this entire category is moving fast.

What Developers Are Saying After Testing It

Vendor benchmarks are one thing. Hands-on reports from people actually running the model are another, and the reaction across Hacker News and Reddit's AI communities has been genuinely mixed — not the uncritical hype cycle you sometimes see with a splashy launch.

The cost enthusiasm is real. Several developers described how little their usage bills moved even under heavy agent workloads, with one estimating the model runs roughly 60 times cheaper than Claude Opus once typical cache-hit rates are factored into a realistic agentic workflow.

But the quality anecdotes are split. One widely shared comparison described a coding task where DeepSeek V4 Pro finished in about 12 minutes for roughly 12 cents but shipped a bug, while Grok 4.6 finished in about 3 minutes for $1.41 with no bug. Another developer running a complex Docker setup said V4 Pro had a few issues where a competing model had none. Neither report is a controlled test, but taken together they're a useful reality check against the headline benchmark numbers.

The most repeated technical critique was methodological: several commenters pointed out that a single trial on a non-deterministic system shouldn't move anyone's opinion much, and that the evaluation harness itself can matter as much as the underlying model. That's a fair and recurring point across the AI benchmarking world generally, not unique to DeepSeek.

The most repeated feature complaint was the lack of vision support — DeepSeek has reportedly deprioritized image understanding on its roadmap, and it came up repeatedly as the single biggest gap for developers who wanted to use V4 Pro for multimodal agent work.

The most repeated caution was about data privacy. DeepSeek's privacy policy permits the company to train on user prompts and completions, and this was flagged more than once as a real consideration — not a hypothetical one — for any business planning to route proprietary or customer data through the API.

One Reddit comment summed up the benchmark jump well, noting the DeepSWE score moved from roughly 7 to nearly 63 across two consecutive releases — a genuinely unusual pace of improvement worth watching, even before independent verification catches up.

Pros and Cons

Pros Cons
Dramatically cheaper than most frontier-class competitors, even after the price hike Pricing has changed twice in under four months — budgeting is harder than it looks
Large agent-benchmark gains over its own preview version All benchmark claims are vendor-reported; no independent verification yet
1M-token context window and 384K max output 0813 model weights not yet published openly
Drop-in compatibility with OpenAI Responses API and Anthropic-style endpoints No vision/image input support, and it's not on the near-term roadmap
Existing integrations keep working with zero code changes Privacy policy allows training on prompts and completions

The Bigger Picture: DeepSeek's Funding Race


It's worth zooming out from the model itself for a moment. DeepSeek's business trajectory has been unusual by any AI industry standard. The company went from a relatively low-profile hedge-fund side project to a global name literally overnight in January 2025, when DeepSeek-R1 briefly became the most downloaded free app on the U.S. iOS App Store and reportedly contributed to a sharp single-day drop in Nvidia's share price.

For most of its existence since then, DeepSeek deliberately avoided outside investors — an unusual stance in an industry where nearly every major lab has raised enormous capital rounds. That changed in June 2026, when the company took roughly $7.4 billion in its first outside financing round. By July, Reuters was already reporting DeepSeek was in talks for a follow-up round that could value the company near $74 billion, with some reporting pegging the target raise around $8 billion.

DeepSeek has said it wants to roughly double staffing across departments, including its data-center and AI-agent teams, and has reportedly stepped up private hiring of chip-design engineers — a move that could, over time, reduce its dependence on chip suppliers like Nvidia and Huawei. Training and running frontier-scale models is capital-intensive in a way that's easy to underestimate from the outside: data centers, specialized chips, and senior AI talent are all in short supply and high demand globally.

Seen through that lens, this week's price increase isn't really a standalone pricing decision. It's one visible piece of a company trying to convert viral popularity into a durable, well-capitalized business — the same challenge every fast-growing AI lab eventually has to solve.

What This Means for North American Businesses and Developers

If you're a developer, startup, or IT decision-maker in the U.S. or Canada evaluating your AI model stack, here's what actually matters from this release:

  • If you're already building on DeepSeek's API, your integration keeps working, but you should re-check your cost model before August 16–17, when the new peak/off-peak pricing lands. Workloads that can shift to off-peak Beijing hours will save meaningfully.
  • If you're evaluating DeepSeek for the first time, treat the benchmark claims as promising but unverified. Run your own evaluation on your actual use case rather than trusting a single vendor chart or a viral social post.
  • If your workflows involve sensitive or proprietary data, read DeepSeek's privacy terms carefully before routing that data through the API, given the training-on-prompts policy.
  • If image or document-vision input is core to your product, V4 Pro currently isn't the right fit — that gap remains unaddressed.
  • If cost is your primary constraint and your tasks are text- and code-based agent workflows, DeepSeek remains one of the most inexpensive frontier-class options on the market, even post-hike.

Should You Actually Switch to DeepSeek V4 Pro?

There's no universal answer here, but a few practical guardrails can help:

Run a small pilot first. Point a non-critical slice of your agent workload at V4 Pro 0813 and compare cost, latency, and error rate directly against whatever you're using today. The Hacker News anecdotes above show real variance in quality, so your own numbers matter more than anyone else's benchmark chart.

Pin your model version. Because the API name didn't change even though the underlying model did, make sure your monitoring and evaluation harness are tracking the actual build (0813) rather than assuming stability just because the endpoint string looks the same.

Model the post-hike costs, not the launch-day costs. Whatever number you calculate today, rerun it against the new peak/off-peak schedule before you commit to a production migration.

Don't treat it as a Claude or GPT replacement yet. Until independent benchmarks confirm DeepSeek's numbers, it's more accurate to think of V4 Pro 0813 as a very strong, very cheap challenger with an unproven track record outside its own test suite — not a settled substitute for established frontier models in production-critical use.

What's Next

A few things to watch over the coming weeks: whether independent benchmark labs publish results that confirm or complicate DeepSeek's own numbers; whether the company follows through on releasing open weights for the 0813 build, as it has for prior releases; how the new funding round at a reported $74 billion valuation progresses; and whether rival Chinese labs like Moonshot AI, Alibaba, and ByteDance respond with their own agent-focused releases in the coming months, given how competitive that domestic market has become.

Frequently Asked Questions

Is DeepSeek V4 Pro officially released now?

Yes. DeepSeek moved V4 Pro out of preview status on August 13, 2026 (listed August 12 on OpenRouter), with the production build labeled DeepSeek-V4-Pro-0813, replacing the April 24, 2026 preview version.

How much does DeepSeek V4 Pro cost right now?

At launch, pricing stayed at preview-era rates: $0.435 per million input tokens on a cache miss, $0.003625 per million on a cache hit, and $0.87 per million output tokens. A new peak/off-peak pricing schedule with increases of roughly 50% to 1,100% takes effect days after launch.

When does the DeepSeek V4 Pro price increase take effect?

The new peak and off-peak pricing schedule begins at 00:00 Beijing time, which various outlets have reported as landing between August 16 and 17, 2026 depending on time zone conversion.

Is DeepSeek V4 Pro better than Claude or GPT?

On DeepSeek's own vendor-reported benchmarks, V4 Pro 0813 scores close to Anthropic's Fable 5 and ahead of Claude Opus 4.8 on several agent-coding tests. No independent lab has reproduced these numbers yet, so treat them as unverified until third-party testing catches up.

Are the DeepSeek V4 Pro benchmark scores independently verified?

No. As of publication, the reported gains were shared through DeepSeek's own channels, and no third-party lab had published a matching independent test run.

Are the DeepSeek V4 Pro 0813 model weights open source?

Not yet. DeepSeek has a history of eventually releasing open weights, but the 0813 checkpoint's weights had not been published as of this article, even though related agent tooling was open-sourced.

Why is DeepSeek raising prices right after launching V4 Pro?

DeepSeek is reportedly pursuing a new funding round near a $74 billion valuation after raising about $7.4 billion in June 2026, and is trying to roughly double staffing. Higher, time-based API pricing points to a shift toward covering the real infrastructure cost of running frontier models at scale.

Does DeepSeek V4 Pro support image or vision input?

No. Vision input isn't part of V4 Pro 0813, and it's the most requested missing feature among developers who have tested the model, with no announced roadmap timeline.

Is it safe for U.S. businesses to use DeepSeek's API?

That depends on your data policy needs. DeepSeek's terms allow the company to train on prompts and completions, which developers have flagged as a real consideration for businesses handling sensitive or proprietary data.

Conclusion

DeepSeek's official V4 Pro launch is a genuinely significant moment in the AI agent race — a real post-training upgrade, not a rebrand, that closes most of the gap to top-tier Western models on the company's own tests. But it arrives with two big caveats that shouldn't get lost in the excitement: those benchmark gains are still unverified outside DeepSeek's own walls, and the pricing that made this model famous is about to look a lot less like the disruptive bargain it once was. If you're building with DeepSeek, the smart move right now is the same one that applies to any fast-moving AI release: test it yourself, watch the independent benchmarks as they land, and budget for the price you'll actually be paying next week — not the one advertised on launch day.

Found this breakdown useful? Share it with a developer friend or teammate who's weighing DeepSeek against Claude or GPT for their next project.

Related Reading on Mustrend

Sources

This article was fact-checked against multiple independent sources as of August 13, 2026. Benchmark figures attributed to DeepSeek are vendor-reported and have not been independently verified at time of publication; pricing details are subject to change — always confirm current rates on DeepSeek's official pricing page before making production decisions.