The day I had access to an LLM I could honestly tell to "please read all 1M tokens", I got a little excited and started scanning twelve years of indie-developer notes that had piled up in the drawers of my desk. App pitch decks, AdMob mediation memos, store-review correspondence, draft replies to user reviews — a mixture of paper and text files that I had honestly lost the full picture of myself.
For the past two months I have been throwing something into Gemini 2.5 Pro's long context almost every day. Code reviews, research, organizing operational notes, checking drafts. After two months of doing it, I found the line between "work that 1M is genuinely good for" and "work where short repeated prompts still win" was much sharper than I had expected. This post is an honest record of where that line falls.
The setup for two months
The environment is mundane. I called AI Studio and the API roughly half and half from a MacBook Pro, with contexts ranging from about 50k tokens on light days to ~800k tokens on heavy days. My app business has been running solo for twelve years, so pitch decks and mediation memos are scattered across year-by-year folders. Submission requirements and notes from my parallel art practice are mixed in too. In short, the material was "everything written down by the same person across a long career".
A note on cost: long-context Gemini 2.5 Pro inputs do scale visibly with your token count. Across two months, the ~110 long requests I made added up to a few tens of dollars. Each request feels like "a few dollars to maybe ten-something dollars". Coming from short LLM calls it is a shock at first, but if you compare it to hiring a person to sort the same archive for a day, the magnitude is fine. I am paying it willingly for now.
Where 1M shines, #1: getting one bird's-eye view of a scattered archive
This was the use I valued most.
I packed twelve years of operational notes into about 600k tokens and asked: "Across 2014 to today, how often have I reused the same apology template in user-review replies?" The answer was sharp. It listed the actual phrasings and counts, the per-year trends, and pointed out that "the vocabulary has become stiffer since around 2019".
A short LLM cannot see this kind of cross-cutting trend when you feed it in slices. RAG can pull related fragments and summarize them, but it is bad at narrating the overall flow. With 1M context you can do something closer to "reading" than to "searching".
When you build solo for a long time, daily fixes inevitably bias toward the screen or function in front of you — local optimization. The habit of re-measuring the whole structure is the first thing that slips when you are busy. A long context window turned out to be a quiet but effective tool for taking that bird's-eye measurement again. Even reading the whole pile of notes once a month and asking for trends surfaces my own operational habits.
Where 1M shines, #2: cross-checking large MDX archives
The second practical use was checking MDX archives across the four Lab sites I run, including Gemini Lab.
For instance, I asked, "Among articles published on Claude Lab in the last month, which ones overlap with existing Gemini Lab articles?" by passing both sites' rendered HTML at once. Pro listed the overlaps slug-by-slug, and separated them into "overlapping but different angle, keep both" and "overlapping and thin, candidate for consolidation". That is good enough as a pre-review signal.
This only worked while the input stayed within roughly 300k–500k tokens. Beyond that, accuracy started to slip noticeably — see the next section.
Where 1M shines, #3: design reviews against past-self statements
When I revisit an architectural decision, giving Pro access to past memos speeds up the discussion.
For example: "Three years ago I documented that the AdMob mediation order depended on a specific SDK version. Why?" By handing it both the 2023 memo and the current codebase, it returned only the contradictions: "The 2023 memo states X, but the current code has invalidated that premise." Claude and ChatGPT can do similar things, but Gemini 2.5 Pro feels the most natural at carrying old material into the present without slicing it. The real power of 1M tokens might be less "search" and more "a conversation with your past self".
Where 1M fails, #1: strict, line-level fact extraction
Time to be honest. Some work bit me when I leaned on 1M too hard.
The first failure was: "Read 30 art submission PDFs and produce a unified deadline table." Pro returned a clean-looking table, but on closer inspection 1–2 deadlines were misaligned with the wrong submission. There is a "skim-and-skip" failure mode in long-pass reading, the same kind a human reader makes when going fast.
This kind of task ends up faster and safer if you slice it into per-PDF prompts and verify accuracy file by file. Long context is for bird's-eye view, not for line-strict extraction.
Where 1M fails, #2: small, fast back-and-forth
I was slow to notice the obvious: short questions don't need long state.
When I want "three options for a function name", I keep wanting to ask while the long context is still loaded. Responses are correct, but latency feels 3–4× longer, and pricing follows. I have learned, slowly, that 2.5 Flash, or simply trimming the context before calling Pro, wins in throughput.
Where 1M fails, #3: contexts with heavy confidentiality
This is an operational judgement, not a technical one, but worth writing down. Once you get used to "just dump everything into 1M", you start to feel tempted to upload unannounced app concepts or draft sketches for collaborative artworks.
I restricted that material to the API mode, within what Google Cloud's terms permit, and made a personal rule not to lower the confidentiality bar simply because the tool was convenient. For a long-running solo business, that quiet credibility cost matters more than the small productivity gain.
My current rules of thumb after two months
A short summary of where I have landed.
When I want to "see the whole trend at once", "resume a conversation with my past self", or "tidy a sprawling long text", I spend on 1M without hesitation. It feels like hiring reading time, and the money paid usually feels matched to value.
When the task is "strict numeric extraction", "short repeated calls", or "highly sensitive material", I either trim the context aggressively, switch to Flash, or simply don't send the data at all. Convenient tools tend to teach you that the decision not to use them is where they pay off.
More than the talk of flashy new features, it may be quiet text like this — small records of where I drew the line between using a tool and not — that turns out useful later, to me or to someone else. At least that's what crossed my mind while I was sorting twelve years of notes.
For the next month, my plan is to take the now-organized notes, embed them, and move daily operations onto a short-LLM workflow that doesn't depend on 1M context. My provisional conclusion is that long context is excellent for "the moment of sorting", but not a tool you want to use every single day.
Thank you for reading. If this helps anyone else who has been quietly building things solo for a long time make a better call about when to reach for long context, I'd be glad.