New Models on Perspective AI: How New AI Models Get Added

TL;DR: New flagship models are typically live on Perspective AI within days of release. Perspective AI routes to the labs rather than hosting models, so a launch needs three things: a provider endpoint, a model card, and a credit rate. A new flagship lands inside the one subscription you already hold beside GPT, Claude and Gemini, so it replaces your current default rather than adding a bill.

When a lab ships a new flagship, the practical question is not whether it is impressive. It is whether you can use it today without buying another subscription. On Perspective AI, new flagship models are typically live within days of release, and you can try one on your existing balance the moment it lands. A new flagship does not become a fourth subscription: it appears in the one subscription that already consolidates GPT, Claude, Gemini and Grok, and it replaces your current default only if you decide it should. This page is the standing answer to that question rather than a dated changelog: how a model gets here, why the lag is measured in days, what a launch costs you, and which models never arrive at all. It belongs to our model access section, and the subscription math behind it is in what a month of model access costs.

Is a new AI model available on Perspective AI yet?

Usually within days of its public release. Perspective AI routes requests to the provider that runs a model, so availability starts the moment that provider exposes a public endpoint, not when the model is announced on stage.

That distinction is the whole mechanism, and it is worth stating plainly because it sets both the speed and the limits. Perspective AI does not train models and does not host frontier ones. It curates them, prices them, and puts them behind one picker. The consequence is a catalog that moves at the speed of the labs rather than at the speed of a release cycle, and a small number of models that will never show up no matter how long you wait.

The three things that have to exist first

A model cannot appear in the picker until three artifacts exist. None of them is a training run.

1. A provider endpoint. The lab has to make the model reachable programmatically. This is the step Perspective AI does not control and the one that sets the calendar. Some labs ship an endpoint the same hour as the blog post; others hold it back for weeks behind a waitlist.
2. A model card. Every listing in the catalog carries five fields: what the model is best at, who built it, its context window, what a request costs in credits, and whether it is foundational or open source. A model with no card is a model you would be spending on blind, so the card is written before the model goes live, not after.
3. A credit rate. Model pricing is cost plus a thin margin, so the rate has to be derived from what the provider actually charges. Text models are metered by tokens at a per-model rate; image and video models are metered per generation.

Read in reverse, that list explains the delay. The engineering is not the bottleneck. Waiting for a lab to publish a public endpoint is.

Why the lag is days and not months

Aggregators that build a bespoke integration per model are slow by construction, because every launch is a small project. Routing changes the unit of work. A new model is a configuration, not a feature, so the cost of adding the tenth model in a month is roughly the cost of adding the first.

This is also why the answer to "will you have the next big one" has been yes so far. There is no product decision to make each time. If a lab ships an endpoint and the model is worth using, it goes in, and it goes in for everyone on the same plan rather than as a premium tier. The pattern generalises across labs: the same path carries a GPT flagship, a Claude flagship, a Gemini flagship, a Grok release, and the open-weight families like Qwen, Mistral, GLM, Kimi, and DeepSeek.

The models that never arrive

Honesty about the ceiling makes the rest of the page worth trusting. Three categories do not make it in:

• Models with no public endpoint. If a model is reachable only inside its lab's own app, no aggregator can route to it. This is a policy choice by the lab, and there is no workaround.
• Interface features dressed as models. A lab's document workspace, its voice mode, or its custom-assistant builder are product surfaces, not models. Access to the underlying model does not carry them across, and you should assume it does not.
• Models withdrawn by their provider. A proprietary model can be deprecated by the lab that runs it, and when the endpoint goes, the listing goes. Open-weight models behave differently here, which is the subject of our comparison of open-weight and frontier models.

Does a New Flagship Mean a New Bill, or One Subscription?

Nothing extra. A newly added model draws from the same credit balance as everything else in the catalog, on the same plan, at $14.99/mo Starter or $49.99/mo Pro. There is no per-lab subscription, no upgrade gate on new arrivals, and no separate bill for the model that happens to be trending this week.

The alternative is familiar to anyone who has run a two-subscription stack: a flagship lands, you want to test it against your current default, and testing it properly means paying a second vendor a full month to answer a handful of questions. Consolidation removes that decision. The comparison is a few requests, not a procurement exercise, and our guide to the guide to choosing a model per task covers how to make the comparison a fair one.

Reading a new model card before you spend on it

A new model is the moment the card earns its place, because you have no intuition for the model yet. Three of the five fields do most of the work on day one.

The context window tells you whether the model can hold the thing you want to give it. A launch headline rarely mentions this and it is often where a new model is weakest or strongest relative to the one you already use. The cost line tells you what a request will draw before you send it, which matters more on a new model than an old one because a frontier launch is frequently the most expensive item in the catalog on the day it arrives. The foundational or open source label tells you where the model runs, and therefore how durable its availability is.

Reading those three before the first prompt turns a launch from a novelty into a decision. It is the difference between "the new one is out" and "the new one is worth 4x the credits for this specific task and nothing else."

Testing the new model against your current default

The reason to care about arrival speed at all is that a new flagship is only useful relative to what you were already using. Put both in the same thread, send the same prompt, and read the two answers next to each other. Switching models mid-conversation keeps the context, so the second model sees the same setup as the first rather than a fresh start, and the method is covered in detail in keeping context when you change model.

Most launches survive about three prompts of this treatment before you can say something specific about them. Some become the new default for one task and nothing else. That is a normal and useful outcome, and it is only reachable if trying the model is cheap enough to be casual.

When the next flagship lands

The pattern to expect is unglamorous and repeatable. A lab announces, an endpoint appears, a card and a rate get written, and the model shows up in the picker for everyone already paying. You do not need to watch for it or upgrade for it. Open the picker after any major release and check whether the name is there, then spend three prompts finding out whether it deserves your default slot. The full catalog and how it is organised is on the Perspective AI model catalog page.

FAQ