DeepSeek announced DeepSeek-V3.2-Exp on September 29, 2025, and the headline change is a price cut: API prices dropped by more than 50%, effective immediately. The model itself is an experiment built on V3.1-Terminus, and the mechanism behind the discount is DeepSeek Sparse Attention (DSA) — a fine-grained sparse attention method the company says speeds up training and inference on long context with minimal impact on output quality.
What actually changed
Three things shipped at once. First, the model: V3.2-Exp is now live on the app, web, and API. Second, the pricing: the API got a 50%+ cut with no change to how you access it. Third, the open-source release — weights are on Hugging Face, a tech report explains the DSA approach, and DeepSeek published key GPU kernels in TileLang and CUDA, noting that TileLang is the faster path for research prototyping.
The honest caveat
DeepSeek’s own framing is that DSA has “minimal impact” on quality and that benchmarks show V3.2-Exp performs on par with V3.1-Terminus. That’s a parity claim, not an improvement claim — this release is about cost, not capability. If your workload is long-context heavy, half-price tokens are a real budget change. If it isn’t, the practical difference may be small.
Run your own comparison before Oct 15
DeepSeek kept V3.1-Terminus available on a temporary API until October 15, 2025, 15:59 UTC, specifically so builders can run comparison tests between the two models. That window is the useful part of this announcement. A/B your actual prompts — not a generic benchmark — across both endpoints while the old model is still reachable, and you’ll know whether DSA holds up on your traffic before the comparison path closes.
Why this matters for your cost model
Sparse attention changes the shape of long-context pricing. If the technique holds, context that used to cost quadratically gets cheaper, and that shifts which tasks make economic sense to route to DeepSeek. If you’ve been tracking how DeepSeek’s first model that thinks while calling tools changed agent economics, this is the same ecosystem moving on a different axis — this time the lever is inference efficiency rather than capability.
One limitation to keep in mind: the announcement came from DeepSeek’s API docs, and the supplied material is the publisher’s own release copy, so the parity claim deserves independent verification before you migrate anything latency- or quality-sensitive.
Practical next step: spin up the comparison API, run your top three most expensive workloads against both models, and decide before October 15. If DSA holds, the discount is free savings; if it doesn’t, you’ll have data instead of a guess.