Gemini has been the default tab on my phone, my laptop and my browser sidebar for the past four weeks. Not as a novelty. As the thing I actually opened when I needed a second opinion on a spreadsheet formula, a fix for a Python script that broke at 11pm, or a summary of a 90-page report I had no intention of reading line by line. Roughly 200 prompts later, here is what held up, what annoyed me, and where the money is or isn’t well spent.
How I put it through its paces
Three entry points: the Gemini app on a Pixel, the side panel inside Chrome, and the API using the Flash model for bulk jobs. Most of the testing ran on Gemini Advanced through Google One AI Premium, with one deliberate week on the free tier to see what actually changes. The task mix was deliberately unglamorous: summarising meeting notes, debugging code, drafting emails, reading charts, and answering the sort of half-formed question you’d normally bounce off a colleague in the corridor.
Speed and everyday responsiveness
The first thing you notice is that the Flash models are quick. Short prompts come back in a second or two, fast enough that you stop watching the spinner. Pro is slower, usually somewhere between 8 and 15 seconds on anything with real thinking behind it, but that pause generally earns its keep on difficult problems.
Live voice mode is the sleeper feature. I used it to talk through a database schema while walking the dog, interrupting and correcting mid-sentence, and it kept up. Latency on interruptions is still noticeable, roughly a beat behind a real conversation, but it’s close enough that you forget you’re talking to software.
The annoyance is output pacing. Long answers often arrive in one burst, then the model keeps generating for another 20 seconds adding qualifications you never asked for. You learn to stop reading once the substance lands.
Reasoning and long-context work
I fed it a 380-page technical PDF and asked it to flag contradictions between chapter 4 and chapter 11. It found two candidates. One was a genuine inconsistency in the source material. The other was a false positive caused by the two sections using the same term differently. That’s a decent hit rate for a single prompt with no guidance, and a useful reminder that the model reads fast but doesn’t always read carefully.
Where the large context window actually helps
Dropping in an entire codebase folder, a full quarter of support tickets, or a long contract and asking cross-cutting questions is where Gemini pulls ahead of most rivals. Multi-hop questions like “which of these tickets mention billing and were escalated twice” work well. What still fails is precision counting. Ask how many times a specific phrase appears and you should verify it yourself.
Plain logic is solid, if not spectacular. Unit conversions, date arithmetic, multi-step word problems: all fine. Ambiguous instructions remain the weak spot. It will pick an interpretation and run with it rather than asking which you meant.
Coding performance
Python is the strong suit. A 300-line refactor request came back clean and idiomatic, with the project’s existing naming conventions respected because I pasted two example functions first. Paste examples and Gemini gets dramatically better. Skip that step and it invents its own style.
Debugging is the most useful workflow I found. Paste an error trace plus the offending function and it usually identifies the real cause within one or two attempts, rather than the surface symptom. Large multi-file codebases are harder, purely because you can only feed it slices. It loses the thread on architectural questions that span eight files.
Writing and editing
Earlier versions of Gemini had a house style built on breathless adjectives and three-item lists. That’s largely gone. Ask for 300 words on a dry topic and you get something flat but usable, which is exactly what a first draft should be. Tone control works well when you give it a sample of your own writing to match.
Where it slips is structure. Left alone, it defaults to a thesis-plus-three-points shape regardless of whether the material suits it. If you want a comparison of subscription tiers to read like a real buyer’s guide rather than a spec sheet, you have to ask for narrative and stakes. The ChatGPT free vs paid breakdown is a good example of the shape most people actually want from that kind of piece.
Images, audio and screenshots
Screenshot debugging is quietly excellent. I sent a photo of an error dialog, got a correct diagnosis and a fix in one reply. Chart reading is mostly reliable, though it mislabelled a legend once and reported the wrong series, so verify anything numeric before it goes into a deck.
Audio handling has improved a lot. A 40-minute recording with two overlapping speakers got transcribed and separated correctly, including a few technical terms I expected it to mangle. The wider Google ecosystem is leaning on this too, including the coaching features in Google’s new Fitbit Air fitness tracker, which leans on the same underlying models to interpret workout data.
Inside Google’s own apps
Gmail thread summaries, Docs drafting, and the Android assistant replacement are where Gemini has an advantage no rival can copy quickly. Asking it to “find the invoice from the plumbing company in March and summarise what was charged” works because it can see the mailbox. That’s a genuinely different product from a chatbot in a browser tab.
On phones the difference shows up in system-level tasks, and how well that integrates varies by hardware. The Pixel 11 Pro XL versus Galaxy S26 Ultra comparison is a decent read on why the same assistant feels different across two $1,300 handsets.
Where Gemini still frustrates
Four weeks of daily use surfaced a consistent set of problems.
- Confident wrong citations. It will attach a plausible-looking source to a claim that source doesn’t support. Verify anything you plan to publish.
- Occasional over-caution. Medical or legal-adjacent questions can trigger a refusal when a straightforward factual answer was clearly what you needed.
- Agreeableness. Push back on a correct answer and it sometimes folds and apologises instead of holding its ground. Ask it to argue the opposite case to test whether the original answer was solid.
- Inconsistency across sessions. The same prompt in two different chats can produce noticeably different quality, which makes it hard to build reliable workflows without pinned instructions.
- Lazy answers late in long chats. Once a conversation gets long, output quality thins out. Starting fresh with a summary beats pushing through.
What it costs and who should pay
The free tier is generous enough that most casual users never need to upgrade. Google One AI Premium at around $20 a month buys you the stronger Pro model, higher usage limits, deeper Workspace integration and more storage. If you’re living in Gmail and Docs all day, the integration alone justifies it. If you just want a chatbot for occasional questions, save the money.
The competitive picture matters here. Rivals are aggressive and uneven in quality, and some of the loudest entrants in the space are still struggling to earn their hype, as this take on why Grok hasn’t landed with mainstream users points out. Gemini’s edge isn’t raw intelligence. It’s that the assistant already sits inside the tools you open every morning.
Getting more out of it
The single biggest upgrade to output quality isn’t a setting, it’s context. Paste two examples of the format you want and Gemini’s hit rate roughly doubles. Ask for a draft first, then a critique of that draft, then a rewrite. The second pass is consistently better than trying to specify perfection in one prompt.
Use Flash for anything fast and disposable, Pro for anything you’ll actually send. Save repeatable prompts as Gems so you’re not rebuilding instructions every morning. And check numbers, dates and citations before they leave your screen, because the model writes with the same confidence whether it’s right or wrong. Treated as a fast, well-read colleague who occasionally bluffs, it earns its place in the workflow.