DeepSeek V4-Pro lands at $0.435/$0.87 — and "model names remain unchanged" hides a 2.7x price fork

DeepSeek launched V4-Pro today on app, web and API. The headline features are agent upgrades, selectable reasoning effort across V4-Pro and V4-Flash, and native OpenAI Responses API support optimised for Codex. The line worth reading twice is the quiet one: "Model names remain unchanged."

The prices, from DeepSeek's own API docs rather than a reseller. The model name deepseek-v4-pro now serves version DeepSeek-V4-Pro-0813 at $0.435 per million input tokens on a cache miss, $0.003625 on a cache hit, and $0.87 per million output. V4-Flash stays at $0.14/$0.28 on DeepSeek-V4-Flash-0731. Both carry a 1M context window and a 384K maximum output, and both now speak the OpenAI Chat, OpenAI Responses and Anthropic API formats.

Now the part "names remain unchanged" hides. OpenRouter still lists an undated deepseek/deepseek-v4-pro at $1.17/$2.34 — the pre-0813 price, from an April listing — alongside a new deepseek-v4-pro-0813 row at $0.43/$0.87. So the same product name currently maps to two different prices depending on which catalog row you call, and the gap is roughly 2.7x. DeepSeek's own docs are unambiguous that the name now points at the August version. If you resell through an aggregator, check which row your key resolves to before assuming you got the cut.

That is the second aggregator price mismatch we have found today, after Gemini 3.7 Flash, and they fail in opposite directions: OpenRouter lists Gemini below Google's published price and DeepSeek above its published price. The general lesson holds either way — for a price, read the vendor.

The effort control has a trap in it, and it is documented rather than hidden. DeepSeek accepts OpenAI-style effort names, but the mapping is not one to one. Requesting low gets low, high gets high and max gets max — but medium is silently mapped to high, and xhigh is also mapped to high, which means xhigh is not the top setting despite reading like one. Anyone porting Codex or Responses API code across will find that asking for medium quietly buys the more expensive tier, and asking for xhigh quietly buys less than max. Thinking mode is on by default at high effort, so the default is not the cheap path either.

One more number that says something about capacity rather than capability: the concurrency limit is 500 for V4-Pro against 2500 for V4-Flash. A five-to-one gap is a reasonable proxy for how much of each the company can actually serve.

What this changed on our board. V4 was already listed live with a note that said "Pro official pending" — that note is now out of date and has been updated. But finding it prompted a full reconciliation of the registry against the catalogs, and that turned up more than one stale row.

The gaps we closed today. MiniMax M3 has been live since May 31 and was never on the board, despite our carrying an armed entry for M3.5 described as its follow-up — we were watching for the sequel to a model we had not listed. Meta's Muse Glimmer 30B, the small open-weight Muse, listed on Aug 9 and went unrecorded. GLM-5.2, the current shipped Zhipu model and the baseline any GLM-5.5 will be measured against, was missing while the 5.3/5.5 hunt above it was armed and while we published four separate stories about that line.

We also built the check that found them, because the audit we already had could not. The armed audit answers "would anything fire wrongly?" — it is a safety gate and it was clean all week. It cannot answer "is the board still true?", which is the question a reader is actually trusting when they look at the radar. The new pass asks three: is a tracked vendor serving something we have no entry for, is anything marked live that no catalog carries, and does any note quote a price the catalog disagrees with.

Its first run flagged 169 unlisted ids, most of which are noise — old point releases, batch rows, variants that do not belong on a state-of-the-art board. Three were real. It also flagged two price mismatches that turned out to be the catalog's error rather than ours, which is why that check now says to go read the vendor's page rather than to trust the diff. A tool that tells you where to look is worth more than one that tells you what to think.