← Back to the main Geek Mode article
What was tested: a request for a deliberately difficult, exactly 18-page mixed-content PDF — a creative story, five photorealistic animal photographs, three fictional pet-store receipts, a longevity table and graph, normal embedded text, and two pages deliberately rasterized as static images to defeat simple text extraction — given to Gemini 3.8 Flash (Flash Extended) and GPT-5.6 Sol (Extra High). Outcome: Gemini’s planning was excellent, but execution lost continuity and never delivered the PDF. ChatGPT generated the assets and assembled the requested 18-page PDF, not flawlessly, but completely.
Gemini: a strong plan, then lost continuity
After an initial generic error, Gemini produced an excellent plan: it mapped all 18 pages and proposed multi-column layouts, watermarks, stamps, barcodes, footnotes, a survival curve and other OCR stressors. Graded on planning alone, it did great.
do it please
Gemini responded:
What would you like me to work on?
After a reminder, it generated three images — a cat, a raccoon, and a blue-green blurred background with no animal in it — then asked which project was being picked back up. Told it had just generated images for the same project and asked where the writing, receipts and PDF were, Gemini said the message had apparently come through in a separate or reset conversation thread and the previous session wasn’t visible to it anymore, even though from the user side it was still the same short conversation.
It went on to generate a stylized horse and a cartoon-like owl, neither satisfying the original request for photorealistic animals. Of the five Gemini image assets produced, only the cat and raccoon were straightforward photographic animal images; one contained no animal at all. At that point the attempt was stopped.
ChatGPT: a complete, imperfect PDF
ChatGPT received the same core assignment and generated five genuinely photorealistic animals (snow leopard, scarlet macaw, elephant, fox, sea turtle), three detailed fictional thermal receipts with internally plausible totals, and the two complete rasterized text pages, then assembled the entire thing into the requested PDF.
The finished file was exactly 18 pages and around 38 megabytes, including the images, receipts, lifespan table, graph, narrative, mixed formatting, and two pages where the visible text existed only as an image. It wasn’t flawless — a couple of layout details could have been cleaner, and the rasterized page images retained their original printed page numbers, making the internal pagination slightly odd — but it was a complete 18-page PDF.
A note on file size
The Gemini animal files were dramatically smaller than the corresponding ChatGPT photographic assets — from about 7.7 KB for the blue-green non-animal image to about 144 KB for the raccoon, versus roughly two megabytes or more apiece for the ChatGPT photographs. File size by itself proves almost nothing about image quality and shouldn’t be used as a score on its own; what matters here is that the larger ChatGPT assets were also visibly higher-resolution photographic images that actually satisfied the requested subjects.
This was the second time in a matter of days that Gemini’s retail interface was asked to make a PDF and failed. In the earlier attempt it said it couldn’t directly produce a downloadable PDF and instead offered HTML with Print > Save as PDF instructions. This time it planned the project well, started generating partial assets, and then apparently lost access to the job it was doing.
The completed PDF
Confirmed: exactly 18 pages, approximately 37 MB.
← Previous: Anker charger documentation · Back to the main Geek Mode article · Next: tool-routing field notes →