DeepSeek cuts V4 Pro pricing 75% to $0.435 input and $0.87 output

• DeepSeek said deepseek_ai's announcement makes the 75 percent V4 Pro discount permanent, taking first-party pricing to $0.435 per million input tokens, $0.87 per million output tokens, and about $0.0036 per million cached input according to testingcatalog's pricing breakdown.
• Artificial Analysis said the cut moves V4 Pro onto its cost versus intelligence frontier, and AiBattle_'s summary of the updated index said its estimated cost to run the Intelligence Index fell from $1,071 to $268.
• The new rates are exactly one quarter of the original list price, as AiBattle_'s earlier pricing note flagged before the official post landed.
• Reactions split fast: kimmonismus tied the price move to V4's efficiency claims, while bridgemindai argued the cheaper tokens still do not fix weak value on harder tasks.
• The price reset also lands as kimmonismus citing Bloomberg says DeepSeek is advancing a $10.29 billion financing round, a useful bit of context for how aggressively it can price first-party inference.

You can check DeepSeek's own announcement, browse Artificial Analysis' provider page, and see the weirdest detail in teortaxesTex's timing joke: the "temporary" full price barely had time to exist before the permanent cut replaced it. kimmonismus also pointed back to V4's efficiency claims, while the Bloomberg-citing follow-up added a financing backdrop that makes the pricing move look less isolated.

DeepSeek's announcement is short, but the price sheet is not subtle.

The permanent first-party rates now in circulation are:

• Cached input: about $0.003625 per million tokens, per testingcatalog's pricing breakdown
• Input: $0.435 per million tokens, per testingcatalog's pricing breakdown
• Output: $0.87 per million tokens, per testingcatalog's pricing breakdown

AiBattle_'s earlier pricing note said the post-promotion "official full price" would be one quarter of the original price, which means the discount effectively became the new baseline. teortaxesTex's post focused on the same wrinkle from the outside: the old list price appears to have had a very short shelf life.

The most concrete third-party read came from Artificial Analysis, which recalculated V4 Pro at first-party pricing and said a 7:2:1 cache-input-input-output blend lands around $0.18 per million tokens.

Its pricing math produced three numbers that matter:

• Artificial Analysis' estimated cost to run its Intelligence Index on V4 Pro dropped from about $1,071 to about $268, according to AiBattle_'s summary of the updated index
• Artificial Analysis said that makes V4 Pro roughly 3x cheaper than Gemini 3.1 Pro Preview on the same index run, 12x cheaper than GPT-5.5 xhigh, and 19x cheaper than Claude Opus 4.7
• AiBattle_'s comparison post said V4 Pro Max scores 51.5 on the index versus Sonnet 4.6 Max at 51.7, while costing about $268 versus $4,206

That is the part engineers will keep bookmarked: not that DeepSeek cut price, but that the cut was large enough for outside trackers to redraw the cost-performance map.

The bullish case is that the pricing drop follows real inference efficiency, not a one-off promo.

kimmonismus pointed to the V4 technical paper's efficiency claims, specifically "only 27% compute and only 10% cache compares to v3.2," and argued DeepSeek is now competing on token efficiency. Yuchenj_UW read the same move more cautiously, floating two explanations seen across the ecosystem: major inference optimizations, or materially cheaper hardware.

The available evidence here is directional, not definitive. The tweet pool does not include a fresh DeepSeek engineering writeup explaining what changed between the temporary sale and the permanent rate card.

Cheap tokens are not the same thing as cheap finished work, and that distinction showed up immediately in reactions.

omarsar0 said the lower rates are "the best news for builders" because the model is already impressive on agentic work. bridgemindai made the opposite argument, saying even $0.87 per million output tokens is not compelling if stronger closed models complete hard tasks more reliably.

That split is more useful than the hot takes around it. Price-per-token got dramatically better in one day. Price-per-successful-task is the part the evidence pool still treats as an open fight.

One more data point arrived in the same conversation: kimmonismus citing Bloomberg said DeepSeek is moving ahead with a $10.29 billion financing round and that Liang Wenfeng remains focused on open-source AI models over short-term commercialization.

If that Bloomberg report holds, the permanent V4 Pro cut sits inside a larger capital story, not just a weekend pricing experiment. It also helps explain why DeepSeek can leave V4 Pro and V4 Flash pressing on the same cost frontier that Artificial Analysis tracks across providers.