DeepSeek V4.1 Flash Tested on a Real Board Deck

The second-cheapest model in the test produced the best ad analysis of all six decks, and recommendations a marketing lead could act on the next morning.

By Ed · Published November 5, 2026 · 5 min read

Twenty-five years starting companies, consulting, owning a multi-million-dollar P&L and leading business intelligence for large enterprises.

DeepSeek V4.1 Flash built a 14-slide board review for 30 cents in 8.2 minutes. It found that Meta prospecting stopped covering its own cost from May, with both ad sets below 1.00 in June, which four of the five other models did not. Its four recommendations are sequenced by month. It found 5 of 8 story points in our checklist, third behind Opus and Argon, at a tenth of Argon’s price.

How this was tested. One prompt, “Northvale 2026 year-to-date board review, with an executive summary and recommendations”, run once per model in the Quiriz PowerPoint add-in on November 5, 2026. Northvale is a fictional US apparel company with stories planted in its data: Google and Meta ad exports and a HubSpot wholesale pipeline, through September 15, 2026. Every step ran on the model under test: plan, build (answering each slide’s question from the data), gap analysis, write and critic; see how a deck is built. Three small parts stayed fixed for every run (chart styling, question routing and a grounding check). Cost is what OpenRouter billed. Build time runs from pressing Build to the last slide written, and leaves out the plan, which waits for a person to approve it. One run per model, so treat close results as a tie. Full method on the benchmark page.

The run

Billed cost	$0.30 ($0.20 answers, $0.05 plan, 2 cents writing and critique)
Build time	8.2 minutes
Analysis calls	67 (the most of seven), reading 2.7 million input tokens
Slides	14 content slides
Stories found	5 of 8 (see the checklist)

What it found

The ad collapse, month by month. Meta return fell from 5.66 in January to 1.09 in June. It then drilled into June: Lookalike-1pct returned 0.75 and Lookalike-3pct 0.79, together $385 less than they cost. Exactly the drill-down a board needs.
The best and worst Google campaigns. Brand-Search returned 7.85 on 18.6% of spend, against 1.66 for PMax-Retarget.
Where Meta should spend instead. Retargeting-DPA returned 4.26 against prospecting’s 2.58.
Wholesale by region, with deals created and won side by side and a clear note on which date each is counted by.

DeepSeek V4.1 Flash deck: Meta platform-reported ROAS fell to 1.09 by June — Meta spend and return by month: the slide that set up the drill-down.

DeepSeek V4.1 Flash deck: both June prospecting ad sets ran below 1.00 — The drill-down: both Lookalike ad sets below 1.00 in June.

What it left out

The Southeast shows up as the region with the fewest wins, but the deck does not reach the 20% win rate or the reason behind it: half of the Southeast’s lost deals were lost on price.

Its recommendations

Four, sequenced by month: cut Prospecting-Lookalike in October (June return 0.77 on $1,680); fund Brand-Search (7.85 against PMax’s 1.66); shift Meta budget to Retargeting-DPA (4.26 against 2.58); and re-point Southeast sales coverage. All four act on real findings, with an owner line and a flat budget. Opus 5.5 and Argon did as well here; Opus went one step further by tying the Southeast move to pricing.

DeepSeek V4.1 Flash deck: sequencing: cut prospecting in October, re-point the Southeast by November — DeepSeek V4.1 Flash’s four recommendations, sequenced by month.

The slide notes

Specific and useful: “Prospecting-Lookalike took 67.7% of May spend and 68.1% in June” is the kind of note that makes a chart readable, and its drill-down slides say what the parent slide could not.

Verdict

The best value in the test. For 30 cents it found the most expensive problem in the ad data and said what to do about it. Compare it with GPT-6.1 Sol, which found less at seven times the price, and with GPT-6 Luna, which cost half as much and did not look at returns.

Build a board deck from your own data

Describe the deck in PowerPoint and see the source of every figure. Free to start.

Try Quiriz free →

Frequently asked questions

Is DeepSeek V4.1 Flash good for data analysis?

On this test, very good for the price: it followed ad returns month by month and found the campaign that stopped paying for itself, which four of five other models did not.

How much did the DeepSeek V4.1 Flash deck cost?

30 cents billed through OpenRouter for the whole run, 20 cents of it answering the data questions. Its list price is $0.15 per million input tokens and $0.60 per million output tokens, and caching brought the answers in 54% under list.

Is DeepSeek better than ChatGPT’s models?

Against OpenAI’s GPT-6.1 Sol on the same deck, DeepSeek was 7 times cheaper, found more of the stories in the data (5 of 8 against 3) and made sharper recommendations; Sol was faster. Full comparison: DeepSeek vs GPT-6.1 Sol.

Is it safe to send company data to DeepSeek?

That depends on which provider serves the model and its data terms, not on the model name. We ran it through OpenRouter on demo data. Before sending real company data, read the serving provider’s retention and training terms.