DeepSeek's 1.6T Flagship Went GA. The Weights Stayed in Preview.

TL;DR

DeepSeek's flagship left preview today, and it did so without a blog post, a changelog entry, or a model card. The pricing docs now list DeepSeek-V4-Pro-0813 as the version behind the deepseek-v4-pro endpoint, and OpenRouter states it plainly: "This is the GA release of DeepSeek V4 Pro." What did not ship are the weights. The Hugging Face repo still describes itself as "a preview version of DeepSeek-V4 series," and no 0813 checkpoint exists anywhere in the deepseek-ai org. Developers found the release by reading API documentation, which is how a company ships something it does not especially want discussed.

What actually shipped

The architecture is unchanged from the April preview: a mixture-of-experts model with 1.6 trillion total parameters and 49 billion active per token. Context is one million tokens, maximum output is 384,000, and both thinking and non-thinking modes are supported. Nothing about the shape of the model is new.

What is new is that the preview label came off. If your code calls deepseek-v4-pro, you are already on 0813. There was no version pin to opt into and no migration window, because the endpoint ID did not change. Whatever the post-training did, it did it to your production traffic this morning.

Published rates carried over from preview, and they remain the reason anyone tolerates this:

That 120x gap between a cache hit and a cache miss is worth internalizing before you build anything on this endpoint. Structure your prompts so the stable prefix stays byte-identical and your bill barely moves; shuffle a system prompt on every call and you pay full freight forever.

The weights are the story

DeepSeek built its reputation on shipping frontier-scale checkpoints under an MIT license, and it has been consistent about it. The V4 series went out as an open-weight preview on April 24 with both Pro and Flash downloadable. When the smaller sibling graduated on July 31, the weights graduated with it: DeepSeek-V4-Flash-0731 sits on Hugging Face today, MIT licensed, described on its card as "the official release, superseding the preview version."

The flagship did not get the same treatment.

The practical effect is that the DeepSeek flagship you can download and the DeepSeek flagship you can call are now two different artifacts wearing the same name. It is like a restaurant that publishes its cookbook every year, then keeps cooking from a revised recipe it has not printed. The book on your shelf is still real, still yours, still free. It is just no longer what comes out of the kitchen.

Nobody should read this as the weights being cancelled. DeepSeek has not said they are coming and has not said they are not; there is simply no published timeline for a 0813 checkpoint. Every open-weight commitment DeepSeek made in April still holds for the artifacts it already released. But if your plan for vendor independence was "worst case, we self-host the flagship," the version you can self-host is now four months behind the version you are actually billed for.

No numbers came with it, either

There is no benchmark table for 0813. The change log stops at July 31, where DeepSeek noted that the Flash update left Pro untouched and that "the official release of DeepSeek-V4-Pro will follow soon." Twelve days later, soon arrived, and the change log did not notice.

Be careful with the figures already circulating. Several write-ups pair the 0813 release with scores of 80.6 percent on SWE-bench Verified, 90.1 on GPQA Diamond, and 93.5 on LiveCodeBench. Those are the numbers DeepSeek published for V4-Pro-Max in the April preview report, not measurements of this build, and they were vendor-run to begin with. Treat them as preview-era context until someone publishes a harness run against 0813 specifically. The Hacker News thread, which reached the front page within hours of a release everyone agreed was "quiet," is currently the densest source of hands-on impressions, and those impressions are mixed in exactly the way you would expect from a model that arrived with no documentation.

And the meter is about to change

Sitting directly above that pricing table is a notice DeepSeek posted on August 6: "We plan to raise the overall pricing for DeepSeek API services in the near future, with a significant increase expected." No figure, no date, just a suggestion to plan accordingly. The South China Morning Post reported it as a reversal for the company that spent two years driving token prices down, arriving a week after a lightweight release that pulled in more demand than the platform comfortably serves.

So the sequence for anyone building on DeepSeek this month reads: prices are going up by an unspecified amount on an unspecified date, and the flagship you would fall back to self-hosting is a build behind. Those two facts are individually survivable and jointly worth a planning meeting.

What to actually do

• Pin the version explicitly if your provider exposes it. On OpenRouter that means deepseek/deepseek-v4-pro-0813 rather than a floating alias, so the next silent graduation does not land on you mid-incident.
• Re-run your own evals. The post-training changed and the vendor published nothing about it. Your regression suite is now the only source of truth about whether this build is better for your workload.
• Audit your cache-hit rate before you worry about the hike. At a 120x spread, the difference between a well-structured prompt prefix and a sloppy one dwarfs any plausible price increase.
• Keep Flash-0731 in the plan. It is the most recent DeepSeek build you can genuinely run yourself, it is MIT, and it is the one that does not depend on a pricing page you do not control.

Key Takeaways

• DeepSeek-V4-Pro-0813 is the GA build of DeepSeek's 1.6T-parameter, 49B-active flagship, live as of August 12 behind the unchanged deepseek-v4-pro endpoint.
• It shipped with no blog post, no change log entry, no model card, and no benchmark table. Developers found it in the API docs.
• No 0813 weights have been published. The Hugging Face flagship repo still says "preview," while the smaller Flash GA build did get MIT-licensed weights on July 31.
• Published pricing is unchanged at $0.435 per million input tokens on a cache miss, $0.003625 on a hit, and $0.87 output. The hit-versus-miss gap is 120x.
• A notice posted August 6 warns of a "significant" overall API price increase with no figure and no date attached.
• Benchmark numbers being attributed to 0813 are recycled April preview figures for V4-Pro-Max. Nobody has published an independent run against this build yet.

Sources: DeepSeek API pricing documentation, DeepSeek API change log, DeepSeek-V4 preview announcement, April 24 2026, OpenRouter: DeepSeek V4 Pro 0813, Hugging Face: deepseek-ai/DeepSeek-V4-Pro, Hugging Face: deepseek-ai/DeepSeek-V4-Flash-0731, Unite.AI, South China Morning Post, Hacker News discussion