DeepSeek V4 Pro Goes GA — Mind the Weights Gap

DeepSeek shipped the production version of its flagship model the way it ships everything lately: a one-line changelog late on August 13, Beijing time. The deepseek-v4-pro API alias now resolves to DeepSeek-V4-Pro-0813, ending a nearly four-month preview that started April 24. No keynote, no countdown — just a doc edit and a wave of benchmark screenshots.

The reflexive read is "another dated checkpoint." The correct read is that DeepSeek just made its strongest claim yet to frontier-class agentic performance at commodity prices — while two quieter facts complicate the story developers actually care about: the weights for this build aren't published, and the price you'd budget against has an expiration date.

What 0813 actually is

The architecture hasn't changed since the preview: a mixture-of-experts model with 1.6 trillion total parameters and 49 billion active per token, a 1M-token context window, and up to 384K tokens of output. The V4 series' real innovation is economic. Its compressed-attention scheme cuts single-token inference compute to 27% — and KV cache to 10% — of what DeepSeek's V3.2 generation needed at the million-token setting. That's what makes the pricing possible: $0.435 per million input tokens, $0.87 per million output, and a startling $0.003625 per million on cache hits.

That cache-hit number deserves a second look, because it's the one that matters for agents. An agent loop re-reads its entire history every turn; with prompt caching this cheap, a long-running session with a fat context costs almost nothing on the input side. The economics of "keep everything in context and iterate" — which get uncomfortable fast on closed frontier APIs — are nearly free here.

What did change in 0813 is post-training. DeepSeek's reported numbers show modest movement on knowledge benchmarks and enormous jumps on agent harnesses: Terminal-Bench 2.1 from 72.1 to 87.9, CyberGym from 52.7 to 83.3, and DeepSWE from 12.8 to 62.7 versus the preview. The company claims wins over Claude Opus 4.8 on several agentic coding evals while conceding gaps on Humanity's Last Exam and repo-scale tasks — trailing Gemini 3.1 Pro by about seven points on HLE.

Two caveats before you repeat those numbers. They're vendor-reported and independent verification is still pending. And some headline figures cite V4-Pro-Max, the maximum-reasoning-effort mode — expect your latency and output-token bill to reflect that if you chase them.

Still, the shape of the release is legible: this is not a new base model, it's a heavy RL-on-agent-harnesses pass over the April base. The frontier labs spent 2025 proving that post-training on tool use and long-horizon tasks is where capability gains now live. A 5x jump on DeepSWE from the same pretrained weights is DeepSeek demonstrating it can play that game too.

The asterisk on "open"

Here's the wrinkle. DeepSeek's reputation — and this release's framing as an open-model event — rests on the MIT-licensed weights on Hugging Face, where the V4-Pro repo logged over 1.4 million downloads last month. But that repo holds the April preview. As of the GA announcement, DeepSeek hasn't said whether 0813 weights will be published.

If the value of 0813 is the post-training — and it is — then the self-hosting escape hatch currently gets you a materially weaker model than the API. The 12.8-versus-62.7 DeepSWE gap is the delta between what you can download and what you can rent. DeepSeek has historically published weights for its official releases, from V3 through R1, so the likely outcome is that they land eventually. But teams who justified DeepSeek adoption with "worst case, we self-host the exact model" should notice that the hedge now lags the product by a release, at precisely the moment the release matters.

The timing sharpens the point. On August 6, DeepSeek warned developers of a "significant" price increase coming, citing compute costs and surging demand — no amount, no date. So the $0.87 output price is best understood as promotional: real today, load-bearing for nobody's twelve-month budget. When a vendor pairs a capability jump with a pre-announced hike and an unpublished-weights gap, you're watching the classic open-model funnel narrow in real time.

What to do with it

If you're running agentic coding workloads — SWE-bench-style issue resolution, terminal automation, CI-scale batch fixes — 0813 is worth an immediate bake-off. Adoption is deliberately frictionless: the API speaks both OpenAI and Anthropic SDK formats, thinking mode takes a configurable reasoning_effort, and the alias upgrade means existing deepseek-v4-pro callers are already on the new build. That last part cuts both ways: if you need reproducibility, pin the dated snapshot on OpenRouter rather than riding an alias that Chinese regulatory quirks — resubmission is reportedly triggered by version-number changes, hence the date-suffix convention — will keep silently advancing.

Know the constraints. There's no vision modality, so multimodal pipelines are out. Pro's concurrency cap is 500 requests, which matters for large batch jobs. And for organizations with data-governance rules around Chinese-hosted APIs, the compliant path is self-hosting the MIT weights — which, per the above, currently means the April model, not this one.

My verdict: the capability claims are plausible and the pricing is genuinely disruptive for agent workloads today, but this is a rent-the-model release wearing an open-model reputation. Benchmark it now, route your cost-sensitive agent traffic to it if it wins, and treat two upcoming events as your real decision points: whether 0813 weights hit Hugging Face, and what the post-hike price sheet says. If both land well, DeepSeek has the strongest price-performance story in agentic AI. If neither does, you've learned how much of "open" was a release-cycle artifact.