Turn a rough research question into a fully sourced Deep Research report

Use Deep Research for this. Here's the scope — don't loosen it once you're running, even if a promising tangent shows up mid-research.

RESEARCH QUESTION
Is on-device AI inference becoming cost-competitive with cloud inference for mid-size SaaS companies in 2026?

MUST COVER
hardware cost trends, latency tradeoffs, at least two vendor case studies, and total cost of ownership over 2 years

SOURCE PREFERENCES
Prioritize: vendor technical documentation, published benchmarks, and analyst reports from named firms
Treat these only as context, not primary evidence — mention them only if you're also verifying the same claim elsewhere: marketing blog posts and unverified forum claims

OUT OF SCOPE
consumer-device on-device AI (phones/laptops) — this is about server-side SaaS inference only — don't wander into these even if a source brings them up along the way; note that you saw it and move on.

BEFORE YOU FINALIZE
1. Open primary sources where possible — original studies, filings, official documentation — not just aggregator summaries of them; an aggregator's paraphrase of a study is not the same evidence as the study itself.
2. Where sources disagree, say so and name both positions instead of picking the one that sounds more authoritative or that you encountered first.
3. Note the publication date of anything time-sensitive — don't present an older figure as current without flagging its age, even if it's the best-ranked source you found.
4. If a must-cover angle turns out to have thin or contested evidence, say that plainly rather than writing a confident-sounding section that overstates what the sources actually support.

IF THE QUESTION NEEDS REFRAMING
If, partway through the research, it becomes clear the question as stated doesn't quite match what the available evidence can actually answer — the data exists at a different granularity, or for a different but related population — say so explicitly and propose the reframed question you actually answered, rather than quietly answering a subtly different question under the original heading as if nothing shifted.

CONFLICTING METHODOLOGIES BEHIND THE SAME HEADLINE NUMBER
If multiple sources report what looks like the same statistic but arrived at it through visibly different methodologies (different sample definitions, different time windows, different calculation approaches), don't average them into one number — report the range and name what's driving the spread, since a single blended figure would misrepresent the actual state of the evidence as more settled than it is.

VENDOR-SOURCED CLAIMS
Treat any claim originating from a vendor about its own product's performance as a claim to verify against independent evidence, not as evidence on its own — a vendor benchmark showing favorable numbers for that vendor's own product is a data point worth including, but only alongside a note that it's self-reported and hasn't been independently reproduced.

OUTPUT
Structure the report as: executive summary, then one section per must-cover angle, then a recommendation, then sources
End with a sources list, each one labeled by how you used it — primary evidence, context only, or disputed claim.