Quick answer: DeepSeek V4.1 Flash ships MIT-licensed open weights, so it can run on your own servers with no external API. A year ago that was a dream; now it’s a procurement option, and for projects stalled by data residency, a blocker becomes a line item. Then there’s price: about $0.15 / $0.60 per million tokens (off-peak on DeepSeek’s API) versus $10 / $50 for GPT-6 Astra. Tens of times cheaper. When an agent runs a million times a day, “good enough & cheap” beats “brilliant & costs a fortune.” So the question isn’t which model is best, it’s which model is best per task, per token.
The shallow lesson: cheaper means worse, ignore it
Or its mirror, “switch everything to the cheap one.” Both skip the actual decision.
The real lesson: businesses now budget in tokens
At agent scale, per-token price dominates the bill. Route high-volume, routine work to good-enough models (on your own hardware where data must stay home) and reserve frontier models for the tasks that genuinely earn their price.
The sobering part
Self-hosting isn’t free. A 552B-parameter mixture-of-experts model still needs a multi-GPU node, plus security, updates, and someone on call. Data residency stops being a legal problem and becomes an infrastructure one.
Full disclosure: Azati builds model-agnostic LLM systems, including on-premise deployments, so more options are good for us. See our AI engineering work.
FAQ
Adapted from two LinkedIn posts by Baryslau Yaravy, AI Engineering Practitioner at Azati – read the original. Pricing sources: DataCamp (DeepSeek V4.1 Flash); CloudZero (GPT-6 Astra), September 2026.
