Give Grok's DeepSearch a research brief it can't shortcut

You are running Grok's DeepSearch (or Deeper Search) mode on a research question that genuinely needs multiple search-and-read cycles, not a single query it can answer from the first page it opens. Treat this as an agentic research task with sub-questions you expect to see addressed individually, not one broad prompt hoping the agent figures out the right decomposition on its own.

RESEARCH QUESTION
Is on-device AI inference actually cheaper than cloud API calls at the volume a mid-size app would run, or is that mostly marketing framing?

SUB-QUESTIONS TO ANSWER SEPARATELY
One, what does on-device inference actually cost in hardware and battery terms per call at scale. Two, what do current cloud API providers charge per million tokens for a comparable model class. Three, where does the crossover point sit, and does it depend heavily on request volume.

SOURCE DIVERSITY REQUIREMENT
No more than one citation from any single AI lab's own blog or marketing page — every cost claim from a vendor needs an independent source checking it.

WHAT WOULD MAKE THIS ANSWER USELESS
An answer that just repeats one hardware vendor's press release framing without an independent cost comparison would be useless — we already have that press release.

DEPTH VS BREADTH PREFERENCE
Depth over breadth — this feeds a real infrastructure decision, so the crossover-point sub-question deserves the most rigor even if the other two end up shorter.

RESEARCH RULES
Work through each sub-question above as its own search-and-read cycle, not as a single combined query — if the sub-questions are genuinely separable, searching them separately is what surfaces sources that answer one well and would be missed entirely by a blended query built to cover all of them at once. Meet the source-diversity requirement literally: if it specifies avoiding single-domain reliance, do not let three of your five citations come from the same publication even if that publication happens to have written the most convenient-to-cite piece. Prefer a primary source — the original filing, the paper itself, the transcript, the dataset — over a secondary summary of it whenever both are findable, and note explicitly when you had to settle for a secondary source because the primary one wasn't accessible or didn't exist. When two sources disagree on a factual point, do not silently pick the one that sounds more authoritative — state the disagreement, name both sources, and give your read on which is more likely correct and why, or say the disagreement is genuinely unresolved if it is. Watch for the specific failure condition named above throughout the research, not just at the end — if you notice partway through that you're heading toward exactly the kind of answer that was flagged as useless, change direction rather than finishing the shallow pass and noting the problem in a caveat at the bottom. Respect the stated depth-versus-breadth preference: if the ask is for depth on a narrower question, do not pad the answer with tangentially related points just to look thorough, and if the ask is for breadth, do not sink disproportionate effort into one sub-question at the expense of leaving others thin.

OUTPUT FORMAT
1. A direct answer to the core research question in two to three sentences.
2. Each sub-question, answered individually with its own citations — not folded into one paragraph.
3. Any disagreement found between sources, stated explicitly with both positions named.
4. A source list noting which sources are primary versus secondary, and confirming the diversity requirement was met or explaining why it couldn't be.
5. Anything you weren't able to verify to the standard this brief asked for, named specifically.