What AI builders are saying about DeepSeek

Get every new DeepSeek video summarised in your inbox.

DeepSeek as a cost-effective workhorse in multi-model workflows

Across several videos, Jack Roberts consistently positions DeepSeek as a practical, budget-friendly option suited to high-volume and research-heavy tasks rather than flagship creative or strategic work. In his breakdown of Hermes Agent, he explicitly recommends assigning DeepSeek to cheap research tasks while reserving more capable — and more expensive — models for harder reasoning work. This framing is reinforced in his cost comparison, where DeepSeek comes in at roughly $1.30 per token bundle versus $45 for a top frontier model, a gap he uses to argue that intelligent model routing matters far more than defaulting to the most powerful option.

The practical upshot, as Roberts demonstrates, is that DeepSeek earns its place not as a standalone solution but as one member of a coordinated team. In his multi-agent orchestration tests, pairing a capable orchestrating model with DeepSeek as a sub-agent for copy improvement produced results competitive with expensive solo flagship runs. The key caveat he surfaces is that a cheap swarm without an intelligent orchestrator performs poorly — DeepSeek and similar budget models need a strong conductor to deliver value.

Chinese open-weight models gaining real traction with production teams

Both Jack Roberts and Matt Wolfe note, from different angles, that Chinese-origin models are moving beyond hobbyist experimentation into serious production use. Wolfe reports that notable companies are already routing workloads to Chinese open-weight models — citing Lindy's adoption of DeepSeek V4 specifically — largely on the grounds that these models are cheaper, more controllable, and not exposed to certain regulatory risks. He frames this as a broader shift rather than a one-off curiosity, with multiple firms independently reaching the same conclusion.

Roberts echoes this by building DeepSeek V4 Pro directly into his Ministry of Agents orchestration layer alongside other models, routing tasks through OpenRouter with prompt caching to cut token costs. Neither creator makes strong quality claims for DeepSeek relative to frontier models; the tone across both channels is neutral-to-pragmatic, treating it as a capable, economical component in a larger system rather than a model that dominates on capability benchmarks.

Prompting strategy and model routing outweigh raw model power

A thread running through Jack Roberts' coverage is that the gap between expensive flagship models and cheaper alternatives — including DeepSeek — is smaller in practice than pricing suggests, and that how you deploy models matters more than which model you choose. His head-to-head testing found that a carefully orchestrated team of cheaper models could match or closely approach the output of a solo frontier model, with output differences often too subtle for most users to notice. He recommends reserving top-tier models for taste-sensitive or irreversible decisions and delegating volume work to cheaper options.

This philosophy is further embedded in his Hermes Agent tutorial, where he describes assigning specific models to specific task types via OpenRouter — using DeepSeek for cheap research, for instance — as a core skill for moving from a casual user to an effective AI builder. The implication across both videos is that builders who treat all tasks as equally demanding of the best model are leaving significant cost savings on the table.