DeepSeek V4 Pro 0813 is live, at 3x the Flash price

Nothing in your code has to change, again. On 12 August DeepSeek pointed deepseek-v4-pro at a new build called DeepSeek-V4-Pro-0813, same endpoint, same 0.435 dollars in and 0.87 out, new weights underneath. We found out the way everyone else did, by noticing a table on the pricing page had changed, because there's no change log entry and no blog post. Flash got the same treatment on 31 July, and at least that one came with a note. The swap itself is routine by now. What's worth your attention is what it does to the choice between the two DeepSeek models, because for the past two weeks the cheap one had been beating the expensive one, and that's over.

What actually changed

We compared the models and pricing page against its own archived copy. On 9 August the model version column read DeepSeek-V4-Pro. Today it reads DeepSeek-V4-Pro-0813. That’s it. That’s the announcement.

One other row moved with it, and it’s the one developers were waiting on. The Responses API was flash-only until now, with a footnote promising deepseek-v4-pro support in early August. That footnote is gone and the row is a tick. So if you were holding a Responses API migration because your hard calls go to Pro, you can stop holding.

Everything else on the page held still. Same 1M context, same 384K ceiling on output, same prices to four decimal places, same concurrency limits. The change log still ends at 31 July, which was the Flash update. Nothing about 0813.

The weights broke the pattern

Here’s the bit we didn’t expect. Every V4 release so far shipped with weights. The April preview put both Pro and Flash on Hugging Face under MIT on day one. Flash 0731 did the same, ungated, with a vLLM recipe in the model card.

There is no DeepSeek-V4-Pro-0813 repository. We checked the Hugging Face API directly rather than trusting a search box, and the newest thing on the deepseek-ai account is still Flash 0731, uploaded 1 August. So the weights you can download today are the April preview. Not what the API is serving.

Maybe they land next week and this paragraph ages badly. I’d guess they do, honestly, because opening the weights is most of DeepSeek’s leverage and they’ve never skipped it. But if you self-host V4-Pro and you’ve been telling people it matches the hosted model, that stopped being true on 12 August, and nobody sent you a note.

Pro takes the lead back from Flash

Two weeks ago the awkward fact about DeepSeek’s lineup was that the small model beat the big one. Flash 0731 outscored the V4-Pro preview on every agent benchmark DeepSeek published, at a third of the price, which made Pro difficult to recommend for anything.

The numbers going round for 0813 fix that. Terminal Bench 2.1 at 87.9 where the preview managed 72.1. DeepSWE at 62.7 from 12.8. AutomationBench 31.8 from 12.8, DSBench-Hard 67.2 from 31.1. Those come from a DeepSeek chart circulating on X and reported by trade press, and we could not find them on any DeepSeek page. Read them as DeepSeek’s claim about DeepSeek, because that’s what they are.

Put the price next to it and the decision gets simpler, not harder. Pro is $0.435 per million input and $0.87 output. Flash is $0.14 and $0.28. That ratio is exactly 3.107 on both sides, which is a suspiciously tidy number and probably deliberate. You’re paying triple for about five points on a benchmark DeepSeek ran itself.

The concurrency limit is the part people miss. Pro allows 500 concurrent requests, Flash allows 2500. If you’re running a batch job or a fan-out agent, that ceiling will hurt you long before the bill does.

The price warning is still sitting there

Same page, footnote one, and it hasn’t moved since early August:

We plan to raise the overall pricing for DeepSeek API services in the near future, with a significant increase expected. Please plan your usage accordingly. The specific pricing plan will be subject to official notice.

Worth knowing what that footnote replaced. Earlier in August the same slot held a concrete plan: peak and off-peak billing, 2x the regular rate between 09:00 and 12:00 and again 14:00 to 18:00 Beijing time. That’s gone, swapped for a blanket warning with no number and no date. Whether that’s better or worse depends entirely on how much of your traffic was going to land in those windows.

Either way, don’t model a year of spend on today’s rates. DeepSeek has told you in writing that they’re temporary.

Should you switch anything

If your Pro calls are already going to deepseek-v4-pro, you’ve switched. That’s the nature of an in-place swap, and it’s the same trap we wrote about when the legacy aliases retired in July. Re-run whatever evals you trust. The results you have on file describe the April preview.

If you’d moved hard work to Flash because Pro looked bad value, it’s worth another look, though not automatically. Five benchmark points is real on tasks that were failing outright and invisible on tasks that were already passing. Try it on the calls that actually hurt.

And if you self-host, sit tight. The build worth downloading isn’t downloadable yet.

Sources

Model version, pricing, concurrency limits and the price increase notice are from DeepSeek’s own models and pricing page, compared against the archived copy of 9 August 2026 for the diff. The GA description, the endpoint creation time of 15:42 UTC on 12 August and the context and output limits are from the OpenRouter listing and its public models API. Weight availability was checked against the deepseek-ai account on Hugging Face. The 0813 benchmark figures are DeepSeek-reported, circulated via a widely shared post on X and covered by Wccftech, and we have not reproduced them. The preview and Flash 0731 scores are from DeepSeek’s published table of 31 July.

Frequently asked questions