Last spring I was paying for three AI subscriptions and still had eleven browser tabs open. One model handled code, another was better with long documents, and a third existed purely because it understood a language the other two fumbled. OpenRouter Chat put all of them behind a single chat window and a single login. What follows is the walkthrough I wish someone had handed me on day one: not what the product is, but what you actually do once you’re inside it.
Before you type anything: pick a starting model
OpenRouter Chat sits on top of a routing layer that exposes more than a hundred models through one interface. If you want the big-picture version of that, the all-in-one OpenRouter Chat explainer covers the architecture and the pricing model. The practical decision is narrower: which name in that dropdown do you pick first?
Message limits and cost differ from model to model, so it pays to sort them into three buckets.
- Free tier models. Good for testing the shape of a prompt. Expect throttling at busy times and shorter context windows.
- Mid-tier generalists. The sensible default for drafting, rewriting, and summarising a few pages.
- Frontier models. Reserve these for multi-step reasoning, awkward code, or anything where a wrong answer costs more than the tokens.
Start free. Send something small and opinionated: “Summarise this paragraph in one sentence, then list three questions it doesn’t answer.” If the reply comes back in exactly that shape, your account, credits, and routing are all working. That’s your setup check, and it takes about forty seconds.
Learn the model switcher before you touch the settings
Most people settle on one model within a week and then treat it like a fixed appliance. That throws away the main advantage here. Switching models mid-thread is the whole point, and it’s where the useful surprises live.
A prompt that exposes differences fast
Try this across three models in three separate conversations: “Write a 60-word product description for an insulated steel water bottle. Don’t use any adjective ending in -y. Mention the lid in the second sentence.”
What you’ll see, roughly:
- One model nails the word count and ignores the lid instruction entirely.
- One follows both rules but sounds like the back of a cereal box.
- One adds an enthusiastic closing line about hydration that you never asked for.
Two minutes of that teaches you more than a week of reading benchmark tables.
What to actually compare
Ignore the vibes and score each reply on four things: did it follow instructions, did it match the format, did it refuse something it shouldn’t have, and how long did the answer take. After a handful of rounds you’ll have a private shortlist with almost nothing in common with the public leaderboards.
Route by task, not by loyalty
Once you know your shortlist, write it down. Mine lives in a sticky note and reads like this:
Code and bulk rewriting goes to DeepSeek. It’s cheap, it’s quick, and it doesn’t get precious about being asked to change things. If you haven’t tried the family, there’s a solid case for why DeepSeek Chat is worth your attention, and the newer DeepSeek V4 Pro build on OpenRouter handles multi-file code noticeably better than the version I used last year.
Long documents go to whichever model has the biggest context window in your price band. Paste the entire contract or transcript rather than a summary of it. Half the value of a large context window is not having to decide what to leave out.
Tone-sensitive writing goes to a frontier model. Cold emails, apologies, anything a client might screenshot.
That three-line map covers about 80% of my working day. The rest is improvisation, which is fine.
A worked example: messy notes into a client update
Here’s the routine that replaced a forty-minute chore. A meeting produced a page of fragments: half-sentences, a budget figure, three decisions, and one unresolved question.
Step one. Paste the raw notes and ask for a structured extract. “List every decision, every open question, and every number mentioned. Bullet points only, no commentary.”
Step two. Open a new chat and change the instruction: “Turn those decisions into a four-paragraph update for a client who wasn’t in the meeting. Neutral tone, no jargon, under 250 words.”
Step three. Switch to a frontier model and ask one narrow question. “Read this update and tell me anything a cautious client would push back on.”
Three prompts, maybe ninety seconds of typing, and the output beats what I used to produce while half-distracted. The model switch in step three is what makes it work. The cheap model writes, the expensive one audits, and you only pay frontier prices for a single paragraph of critique.
Presets and threads do more work than you’d expect
Two features earn their keep once you’re past the toy stage.
Presets save a model plus its system prompt as a reusable starting point. If you write the same kind of email every Monday, build the preset once and skip the setup. It also stops you from accidentally firing a heavyweight model at a task a cheap one handles perfectly well.
Threads keep context separated. Mixing a coding session and a client draft in one conversation produces strange crossover, where the model starts applying your code style rules to the prose. A fresh thread costs nothing and prevents that.
Keeping the bill boring
Credit-based pricing is easy to ignore until it isn’t. Three habits keep it predictable:
- Do exploratory work on free or mid-tier models, then re-run only the final draft on a frontier one.
- Set the length inside the prompt. “In under 150 words” is cheaper than a follow-up asking it to shorten.
- Check the usage page weekly for the first month. You’ll spot the one expensive habit you didn’t know you had.
When a single model is still the right answer
Routing across providers isn’t automatically better. If you’re doing one thing repeatedly, a daily translation or a fixed formatting job, pick the model that does it well and stop shopping. Constant switching has a cost, and it’s paid in attention rather than money.
The real value shows up in the messy middle: projects where you don’t yet know which model fits, tasks that need one voice for drafting and another for checking, and weeks when a new release lands and you want to test it without signing up for subscription number four. New options do keep arriving, and it’s worth knowing what’s coming. The GPT-6 Astra release is a good example of something worth trying the day it appears on the list rather than three months later, once everyone has already formed an opinion for you.
Start with one free model, one real task from your actual week, and the switcher open in a second tab. The rest of the workflow builds itself from there.