English · 6 min read
DeepSeek V4 Pro-0813: what actually changed (and what did not) for business
On 12 August 2026 the deepseek-v4-pro API started serving 0813 weights. Promo pricing is unchanged. Here's what a small business should do — and why this is not a new product launch.
The short answer: DeepSeek swapped the weights behind deepseek-v4-pro. The API id, calling method, and promo price did not change. This is a weight drop, not a GA relaunch. Keep V4 Flash for volume. Use Pro-0813 when the task is hard enough that Flash's misses would cost more than the extra tokens.
Verified 12 August 2026 against DeepSeek's API docs, pricing, and changelog. Vercel's AI Gateway changelog the same day noted updated Pro weights. We do not invent an Artificial Analysis score for 0813.
What actually shipped
The deepseek-v4-pro endpoint now serves DeepSeek-V4-Pro-0813. If you already call that id, you are on the new weights without a code change.
What did not ship:
- A new product name you have to buy.
- A price change. Promo rates remain $0.435 input / $0.87 output per 1 million tokens, with cache-hit input at $0.003625.
- A Pro GA note on DeepSeek's own changelog. The last relevant entry (31 July 2026) still said an official Pro release "will follow soon." Until that page updates, do not brief your team as if Pro "went generally available today."
Flash is unchanged: $0.14 / $0.28 per 1M on deepseek-v4-flash. That is still the volume pick.
Why the distinction matters
Labs often drop new weights on an existing id. Gateways (Vercel noted this on 12 August) pick them up the same day. Headlines then read like a launch. For a business, the only questions that matter are:
- Did the bill change? No.
- Do I have to migrate? No, if you already call
deepseek-v4-pro. - Did independent benchmarks move? Not on our LiveBench snapshot. We will not invent an Intelligence Index for 0813. An older independent AA figure for a Pro preview is not a score for these weights.
Treat 0813 as "the same cheap hard-reasoning endpoint, newer weights." That is useful. It is not a reason to reopen every model decision in the company.
What to run where
High-volume, behind-the-scenes work — classify enquiries, tag leads, summarise reviews, draft routine replies. Use V4 Flash. A frontier model per call here is waste.
Hard reasoning you still want cheap — long analysis, trickier coding, jobs where Flash's misses show up in review. Use V4 Pro-0813 on the same API. You are paying roughly 3× Flash, not 10× a US flagship.
Everyday chat for the team. Stay on ChatGPT or Claude if that is already the habit. DeepSeek's saving shows up in automation volume, not in replacing a $20 chat seat your people already open.
Frontier API at a sane (not cheapest) price. That slot this week is Grok 4.6 at $2 / $6 — not DeepSeek, and not Grok Bot.
The honest caveats
Changelog lag. Until DeepSeek posts a Pro GA note, do not promise customers or your board that "Pro launched." Say the weights updated.
No new LiveBench row. Our published snapshot is still 25 June 2026. Flash 0731 and the earlier V4 Pro row are on it; 0813 is not a new independent score.
Data and jurisdiction. Same as any DeepSeek call: know where tokens go before you route customer data. The weight drop does not change that review.
The move this week
Do not switch your chat app. Do not treat 0813 as a priced relaunch. If you already use deepseek-v4-pro, you are on the new weights. If you run a lot of tokens, keep Flash as default and pin Pro for the hard slice.
A Business AI Plan ($99 one-time, optional $29/month to keep it current) names that split for your workflow. The free 2-minute AI plan finder is the fastest way to get a first pass.
Sources: DeepSeek API docs, DeepSeek pricing, DeepSeek changelog