DeepSeek reroutes V4-Pro API traffic to a smaller model on September 14

Any service that sends V4-Pro a prompt and hands the reply straight to a parser has four days ^[9]. The price per call falls without anyone doing anything ^[15], and DeepSeek's announcement does not address whether the reply keeps the shape the parser expects ^[7].

At peak rates, output on V4-Pro costs $3.96 per million tokens and V4.1-Flash costs $1.20 ^[15]. That is $2.76 less per million, a cut of 69.7% ^[1]. Off peak, V4.1-Flash is 60 cents per million output tokens and 15 cents per million uncached input, with weekday peak windows doubling both ^[14]. Doubling the 60 cents is where the $1.20 comes from ^[6].

V4.1-Flash holds 552 billion parameters ^[5] and activates 8 billion while it reads a prompt and 16 billion while it writes one ^[6], so under 3% of the model is doing the generating ^[2]. DeepSeek's technical report says key-value cache entries sit in a four-bit floating point format at 890 bytes per token globally, about a quarter of what V4-Flash needed ^[8]. That puts the older model near 3,560 bytes per token ^[3]. Persistent cache on SSDs drops to roughly an eighth of the previous generation ^[9].

The performance case is DeepSeek's own table, run at maximum reasoning effort ^[10]. On the DeepSWE v1.1 software engineering test the model resolved 74.2% of tasks, against 74% for Claude Opus 5 and 62.7% for V4-Pro ^[12]. That is two tenths of a point over Anthropic's model and 11.5 points over the model it is answering for ^[5]. Both American models still lead on GPQA Diamond ^[13]. DeepSeek said tests by multiple parties put V4.1-Flash ahead of V4-Pro on performance, cost, speed and total runtime ^[2], and the account of the launch does not name them ^[8].

Sort the work by who reads the output. A human reader absorbs a model swap. A parser or a schema turns it into a failure rate. Cross that with output token volume and you get four boxes. High-volume calls read by humans collect the full price cut with almost nothing to check. Low-volume calls read by machines need the checking and return little money for it. Test the high-volume machine-read calls first, where the saving is largest and so is the count of calls that can break.

The only version pin on offer is self-hosting. The weights are on Hugging Face under the MIT license ^[16], and DeepSeek said it will work with the open-source community on inference support and explore further deployment options ^[17]. That freezes V4.1-Flash for a team that needs a build it controls. Anyone who shipped image understanding against the experimental model DeepSeek put out in August now has those calls answered by V4.1-Flash, which carries the capability in the model itself ^[7].