Getting Started With Perspective AI: Your First Ten Minutes
TL;DR: Your first ten minutes on Perspective AI: pick a model from the catalog, check what an action costs in credits before sending it, switch models inside the thread if the answer is weak, then clone one agent. Those four steps are what turn the subscription you just consolidated your other AI plans into from an expensive ChatGPT into all AI models in one subscription.
Key Takeaways
- Start with a real task, not a test prompt. The model picker only becomes useful when the answer matters to you.
- Every model carries a card listing what it is best at, who built it, its context window, and what it costs to run, visible before you send anything.
- Credits are dollar-legible and draw from one balance: text metered per token, image and video per generation, agent steps per operation.
- A weak first answer is a routing problem, not a dead end. Switch models inside the same thread and the new model sees the full history.
- Reasoning effort and web search are per-request controls, so the same subscription behaves like a quick chatbot or a slow careful reasoner depending on the request.
- Clone one agent rather than five, connect it to a single channel, and let it earn its credits before you build a fleet.
Quick Answers
What should I do first on Perspective AI?
Open a thread with a task you actually need done, choose a model whose card matches that task instead of accepting whatever is selected, and send it. The habits that make the subscription pay off, reading costs and switching models, only form on real work.
How do I know which model to pick when I start?
Read the model card. Every model in the catalog lists what it is best at, the lab that built it, its context window, and what it costs to run in credits. Match those four lines against the task in front of you. If you would rather not choose, Auto mode picks the model per request.
How do I know what something costs before I send it?
The cost is shown before your balance moves. Text models are metered by tokens sent and received, image and video models per generation, and agent operations individually. The balance reads in dollar terms rather than abstract points.
You just subscribed, so every major lab is now in one subscription and the picker is full. The fastest way to get value out of the first session is to run one real task through four specific steps rather than to explore. This is a setup guide, not a tour. It sits in our use cases section because using the subscription well is itself the use case, and every step below maps to a control you can see on screen right now. If you want the full inventory of what is in the picker first, the Perspective AI model catalog lists the labs and modalities. Otherwise, start here with one task you were going to do anyway. If this subscription replaced a ChatGPT Plus or Claude Pro plan, the first session is also where you find out whether the consolidation actually holds: the test is whether you stop wondering which app a question belongs in.
What should you do in your first ten minutes on Perspective AI?
Open a thread, pick a model that matches the task, read its credit cost on the model card, send one real prompt, switch models if the answer disappoints, then clone one agent.
That sequence is deliberate. Each step teaches one control, and the controls compound: knowing what a model costs is what makes switching feel cheap, and switching cheaply is what makes an agent worth delegating to. Below is what each step looks like in practice.
Minutes one to three: pick the model instead of accepting the default
The single habit that separates a subscription that feels worth keeping from one that feels like a more expensive ChatGPT is deliberate model choice. Every model in the catalog carries a card with four lines on it:
- What it is best at, so you can match the model to the job instead of guessing.
- Who built it, the lab behind the model.
- Its context window, how much conversation and source material it can hold at once.
- What it costs to run, in credits, before you send anything.
Read those four lines once and pick on purpose. If the task is a long document, the context window line is the one that decides it. If it is a quick factual question, the cost line is. Models are labelled either foundational, meaning proprietary frontier models such as the GPT, Claude, Gemini, and Grok flagships, or open source, meaning open-weight families like GLM, Qwen, Mistral, Kimi, and DeepSeek. Perspective AI routes your request to the lab that runs the model. It does not host the frontier models itself, and the label tells you which is which.
If you genuinely do not want to make this decision, you do not have to. Two routing modes ship today: Custom, where you pick the exact model, and Auto, where the platform picks per request. Start in Custom for the first few tasks anyway. You learn more about the catalog in three deliberate choices than in thirty automatic ones.
Minutes three to five: read the price of an action before you send it
Credits are the single unit of usage. One balance covers chat messages, image and video generations, and agent operations, and the balance is displayed in dollar terms rather than as an abstract points score. The important part for a new subscriber is the ordering: the cost appears before your balance moves, not after.
| What you use | How it is metered | What that means on day one |
|---|---|---|
| Text and reasoning models | Tokens, input sent plus output received | A short question costs a fraction of a credit. Pasting a long report costs proportionally more, because what you send counts too |
| Image and video models | Per generation | A flat, known price per image or clip, visible before you generate |
| Agents | Per operation | Each step, tool call, and command is metered on its own, so a preview costs a bounded amount |
Rates are set cost-plus: the real provider cost for that model plus a thin platform margin, refreshed as provider prices move. That is why an efficient small model can cost an order of magnitude less than a frontier reasoning model for the same prompt, and why the gap is worth looking at before you send. The full mechanics, including top-ups and referral credits, are in the explainer on what a credit costs per action.
Two habits fall out of this within the first session. Trim what you paste, because input is billed as well as output. And stop sending easy work to expensive models, because the price difference is on screen rather than hidden in a monthly limit you discover by hitting it.
Minutes five to seven: use the second model you now have in one subscription
Your first answer will sometimes be weak. On a single-lab subscription that is the end of the road: you rephrase, you retry, or you open a different app and paste your context in from scratch. Here it is a routing problem with a one-click fix.
Switching models mid-conversation keeps the full message history. Tap the picker at the top of the thread, choose a different model, and keep typing. The new model reads everything that came before, including the previous model's answer, so "do you agree with the approach above, and what would you do differently" works immediately. You can also combine several models in one thread, letting each answer the part it handles best, or compare two models on the same prompt side by side.
Try it once in the first ten minutes on something small. It is the behaviour the whole subscription is built around, and it does not become instinct until you have done it on a real answer you were unhappy with. The deeper walkthrough, including which other platforms manage it, is in switching AI models mid-conversation.
Minutes seven to ten: clone one agent and give it a channel
An agent is a step past chat: a model plus a persona, memory, and tools, able to act over many steps. Memory here means the agent holds context across a long-running task, so the plan it made in step one still exists in step nine. The lifecycle is three moves.
- Preview. Try a published agent before committing and watch how it actually behaves. Previews are metered per operation, so looking costs a small bounded amount rather than a subscription.
- Clone. Instantiate your own private copy. This is a one-time cost, and what you are paying for is the configuration, hosting, and uptime, not the model weights.
- Run. Every operation the clone performs draws credits, scaling with how much work it does rather than sitting as a flat fee.
Then connect it to one channel. Agents can reach you in Telegram, Slack, Discord, or WhatsApp as well as in the app itself, which is the difference between a dashboard you must remember to open and a contact that messages you when the work is done. Pick the channel where the work already happens, connect a single agent, and give it something genuinely delegable. One working agent teaches more than five half-configured ones. The full lifecycle, including the creator economics that are still being wired, is covered on the page about agents with memory, tools and channels.
Two per-request controls worth learning on day one
Two switches apply across the text catalog and both are per request, not per plan:
- Reasoning effort. Spend more thinking on hard problems and less on simple ones. Turning it up on an architecture decision and down on a reformatting job is the cheapest quality lever available.
- Web search. Let the model answer from live web results when the question needs current information, or keep it off when it does not. Search-native models return answers with citations attached.
A realistic first morning uses both: search on for the question about something that happened this week, effort turned up on the hard decision, then everything off and a cheap open-weight model for tidying up notes. Three separate tools in the single-subscription world, three requests in one thread here.
One more setting worth knowing exists before you need it: private mode keeps a conversation's history in your browser rather than storing it server-side, for the threads you would rather not have synced. Standard mode is the default and syncs across your devices. Details are on the page about browser-local encrypted history.
Where a month of credits actually goes
Starter is $14.99/mo and Pro is $49.99/mo, and each carries a monthly credit allowance. Both plans carry the identical model catalog, because the honest difference between users is volume rather than which models they can reach. Top-ups join the same balance whenever a project month runs long, and referral credits land there too, with no separate redemption rules to learn.
What that allowance buys depends almost entirely on routing. A month spent sending every question to the most expensive frontier reasoning model looks nothing like a month where the cheap models take the easy work and the expensive ones take the decisions that deserve them. That is the whole argument for reading cost lines in minute four, and it is why the next thing worth reading is the task-by-task routing playbook: which model to reach for when the job is code, a long document, cited research, or bulk work you want done cheaply.
FAQ
What should I do first on Perspective AI?
Open a thread with a task you actually need done, choose a model whose card matches that task instead of accepting whatever is selected, and send it. The habits that make the subscription pay off, reading costs and switching models, only form on real work.
How do I know which model to pick when I start?
Read the model card. Every model in the catalog lists what it is best at, the lab that built it, its context window, and what it costs to run in credits. Match those four lines against the task in front of you. If you would rather not choose, Auto mode picks the model per request.
How do I know what something costs before I send it?
The cost is shown before your balance moves. Text models are metered by tokens sent and received, image and video models per generation, and agent operations individually. The balance reads in dollar terms rather than abstract points.
Can I change models without starting over?
Yes. Switching models mid-conversation keeps the full message history, so the new model reads everything that came before it and continues from there. You can also combine several models in one thread or compare two on the same prompt.
Should I set up an agent on day one?
Set up one. Preview a published agent to watch how it behaves, clone it to get your own private copy, then connect it to Telegram, Slack, Discord, or WhatsApp so it reports where you already work. Each operation it runs is metered in credits.
Open the model picker and start with one real task
Everything in this guide happens inside the app: the model card with its cost line, the picker at the top of the thread, and the agent store. Bring a task you were going to do anyway.
Open Perspective AI →