Google’s Gemini has had a strange public life. It turned up in December 2023 with a demo video that got quietly edited, spent most of 2024 being measured against GPT-4, and by 2025 was regularly trading the top spot on leaderboards with whatever OpenAI had shipped that month. Somewhere in there it stopped being “the other chatbot.”
That shift matters more than it sounds. Gemini isn’t just a ChatGPT rival sitting in a browser tab. It’s the engine inside NotebookLM, the assistant buried in Google Workspace, the speech-to-text on Pixel phones, and a developer API that a lot of production apps quietly depend on. So here’s the practical version: what the models actually do well, where they get in your way, and how to squeeze better answers out of them.
One name, several very different models
Ask three people what Gemini is and you’ll get three answers, because Google uses the word for a whole family spanning a phone chip to a data centre. The tiers break down roughly like this:
- Gemini Nano runs on-device, on Pixel phones and inside some Android features. It’s small enough to work offline, which is why live transcription and summarisation can happen without a server round trip.
- Gemini Flash and Flash-Lite are the fast, cheap workhorses. Flash-Lite is built for high-volume work where fractions of a cent per request matter more than depth.
- Gemini Pro is the flagship. Slower, more expensive, and the one worth using for multi-step reasoning, awkward code, or anything where a wrong answer costs you real time.
The other number worth knowing is the context window. Gemini models accept around a million tokens of input, and some Pro variants stretch to two million. That’s roughly 1,500 pages of text in one prompt, or several hours of audio. You won’t need it most days. On the days you do, nothing else in the mainstream feels quite the same.
Where Gemini actually pulls ahead
Long documents that hold together
Dump a 300-page contract, a full compliance manual, or six months of meeting notes into a window and ask a question about page 280. Gemini tends to find it. Models with smaller windows often lose the thread long before that, summarising confidently from whatever survived the truncation. Same story with code: pointing it at an entire repository and asking why a build fails is genuinely useful in a way that pasting one file usually isn’t.
Answers that show their sources
Grounding with Google Search gives you inline citations rather than a confident paragraph with no provenance. For anything time-sensitive — pricing pages, release notes, who currently holds a job — that’s the difference between a tool and a rumour mill. It’s not infallible. It will still sometimes cite a forum post as though it were documentation. But you can at least check the claim in two clicks.
Native video and audio
Gemini was multimodal from day one, not bolted together afterwards. Upload a 40-minute recording and ask for a list of decisions with timestamps. Feed it a screen recording of a bug and ask what went wrong. Ask it to pull the numbers out of a photographed whiteboard. Very few assistants handle that kind of mixed input without you first converting everything to plain text.
The annoyances nobody puts in the launch post
The safety layer can be twitchy on legitimate work. Ask about drug interactions, contract clauses, or security testing and you’ll occasionally get a lecture instead of an answer. It has improved, but the habit survives.
Naming is a mess. Bard became Gemini. Model strings look like gemini-1.5-pro-002. Feature names change between the consumer app, Workspace, and AI Studio, and the same brand label can hide noticeably different behaviour depending on which surface you’re in. If you’re testing something for work, pin the exact model.
Pro-tier latency sags under heavy load, especially mid-afternoon in Europe and North America. Flash exists precisely because of this, and reaching for it by default is often the right call.
What it costs
The consumer app is free, and the free tier is generous enough that plenty of people never pay. AI Studio, which is where you go to tinker with prompts and models directly, is also free with daily limits. Past that, Google One AI Premium bundles Gemini Advanced with 2TB of storage for about $19.99 a month.
The API is a different meter entirely. Flash-Lite sits at a few cents per million input tokens; Pro is an order of magnitude more expensive. If you’re weighing whether a free plan covers your needs, it’s worth reading a straight comparison of what free AI chat tools genuinely give you, because the limits are usually about usage caps rather than capability.
Gemini for developers
The API is available through AI Studio for prototyping and Vertex AI for anything that needs to scale, and it supports function calling, structured JSON output, and context caching. That last one is underrated: if you’re sending the same 50,000-token system prompt on every request, caching the prefix can cut the bill by more than half.
Tooling around it has matured too. Simon Willison’s command-line plugin, which reached version 0.33 of llm-gemini, makes it easy to pipe text into a Gemini model and get output back in a shell script, which is a surprisingly good way to test prompts without building a UI.
Cost discipline matters once you’re past the prototype. Sending every request to the flagship model is the fastest way to burn a budget on tasks a smaller one handles fine, which is why teams increasingly lean on adaptive model routing to keep inference costs down — route the easy 80% of traffic to Flash and reserve Pro for the cases that need it. It’s also good enough at code that people are now vibe-coding entire apps in a single afternoon, which works beautifully right up until the debugging starts.
Small habits that noticeably improve the output
None of these are clever. They just make a measurable difference.
- Put the material first, the question last. Long inputs process better when the instruction arrives after the context, not before it.
- Turn on thinking mode for anything with more than two steps. Extra reasoning time is wasted on simple lookups and genuinely worth it on maths, planning, or debugging.
- Specify the shape of the answer. “Three bullet points, under 20 words each” produces something you can actually use.
- Give it an escape hatch. Telling it to say “not stated” rather than guess cuts hallucinated detail sharply.
- Keep iterating in the same thread. Restarting every time throws away context you already paid for.
Where this is heading
The interesting work has moved past chat. Gemini now powers agentic features that click through websites, fill forms, and complete multi-step tasks on your behalf — uneven in practice, genuinely useful in narrow cases. On-device Nano keeps improving, which means more privacy-friendly features on phones that don’t need to phone home at all. And the release cadence shows no sign of slowing: each generation lands with better reasoning and coding scores than the last, and the gap between the front-runners keeps compressing.
For most people, the deciding factor won’t be a benchmark. It’ll be where the model already sits. Gemini’s advantage is being embedded in the tools you open anyway — your inbox, your documents, your phone’s keyboard. That’s a quieter kind of dominance than a flashy demo, and it’s proving harder to dislodge.