"The Strongest Open Model" Shipped on Its Own Say-So — The Confident Scorecard DeepSeek V4 Pro 0813 Laid Out, and the Verification That's Still Missing
DeepSeek released V4 Pro "0813" as GA. It touts breakthrough pricing of $0.87 per million tokens and top-tier coding scores — but most of that is self-reported, with third-party verification still pending. Here's what developers should confirm before adopting it.
China's DeepSeek released "0813," an update to its large language model "DeepSeek V4 Pro," as generally available (GA) on August 12, 2026. It's listed on the company's own pricing page and can be called through OpenRouter. The pitch: high scores across coding and agentic use cases, paired with pricing that undercuts the field by an order of magnitude. But most of those numbers are, for now, DeepSeek's own tallies, and independent reproduction hasn't yet caught up. From a developer's standpoint, there are a few things worth checking before you take "cheap and strong" at face value.
What actually shipped
V4 Pro is a mixture-of-experts (MoE) model that only activates the parts it needs. Per the published figures, it has 1.6 trillion total parameters but only 4.9 billion activated on each inference pass, and it supports a context length of 1 million tokens. The design carries a large capacity while keeping the cost of actually running it down. 0813 is positioned as a new snapshot with some of those weights swapped in.
The pricing is a steal — but a hike lands on August 16
What draws the eye is the API pricing. The figures below reflect the level at launch, all per 1 million tokens. Note that starting August 16, pricing shifts to a tiered peak/off-peak scheme, with an effective price increase already announced.
| Category | Price (per 1M tokens) |
|---|---|
| Input (cache miss) | $0.435 |
| Input (cache hit) | $0.003625 |
| Output | $0.87 |
When the cache is working, the input side drops to a level you can almost ignore. In agentic setups that reuse long contexts repeatedly, this price gap flows straight into your monthly token bill.
A confident scorecard
The scores DeepSeek puts forward are genuinely high, with figures on coding metrics that stand shoulder to shoulder with top-tier models. Here are the representative ones (all vendor-reported figures).
| Benchmark | Reported score | What it measures |
|---|---|---|
| SWE-bench Verified | 80.6% | Bug fixes in real repositories |
| LiveCodeBench | 93.5% | Competitive code generation |
| Terminal-Bench 2.1 | 87.9 | Multi-step tasks in a terminal |
The 80.6% on SWE-bench Verified is among the very best for an open-weight model, and some tallies put it on par with leading commercial models. DeepSeek also claims improvements of "up to 49.9 points" on agentic metrics.
Don't swallow the "launch-day numbers" whole
This is the crux this time. These scores have not yet been reproduced and verified by an independent third party. Coverage repeatedly carries the caveat that "the benchmark claims await independent verification." On top of that, another review points out that in long, multi-step agentic runs the model tends to break down on "recovering from errors" and "consistency," with roughly 24% of the hardest problems timing out. In other words, the peak performance visible on a static scorecard and the stability to run all the way to the finish in real use don't necessarily line up. It's reasonable to take it in the register of "these concerns have also been raised."
Still, the "run it yourself" room that matters to developers
The V4 line is distributed on Hugging Face as open weights under the MIT license, with self-hosting, fine-tuning, and commercial use permitted (accounts of whether the 0813 snapshot's weights have been released vary by source, so check the distribution page for the latest scope of what's offered before adopting it). The option of keeping your code in-house and bringing only the model down into your own environment has now expanded into a realistic price range — a tailwind for teams leaning toward in-house or on-premises builds. At the same time, running it yourself means shouldering the work of GPU procurement, updates, and evaluation. Cheap API or self-hosting with the weights in hand — which one comes out ahead depends on how sensitive the code you're handling is and how many tokens you're running.
The stance to take now
It's well worth evaluating as a candidate. But whether to adopt it is safest judged not on the reported scores but on results measured against your own real tasks. Rather than a one-shot SWE-bench-style test, run it on multi-step tasks close to your actual workflow, including recovery from failures. Line up the real post-hike pricing against the total cost of self-hosting and compare. Lock down those two before deciding whether to seat it among your existing coding agents — there's no rush.
References: Artificial Analysis: DeepSeek V4 Pro 0813 / OpenRouter: DeepSeek V4 Pro / Tech Times: DeepSeek V4 Pro 0813 Goes GA, Benchmark Claims Await Independent Proof / Morph: DeepSeek V4 Architecture, Benchmarks, Pricing / Codersera: DeepSeek V4-Pro Review
