Models

The model catalog — how models are added, priced and routed.

Perspective AI is model-agnostic by design. Rather than locking you into one lab, the platform curates a catalog of the strongest models available — proprietary frontier models and open-weight models side by side — and lets you move between them freely inside a single interface.

The catalog

The catalog spans the major labs and every modality:

  • Text & reasoning — flagship models from Anthropic (Claude), OpenAI (GPT), Google (Gemini), xAI (Grok), GLM, Qwen, Mistral, Kimi and more.
  • Image — text-to-image and image-editing models.
  • Video — text-to-video and image-to-video models.
  • Search-native — models that answer from live web results with citations.

Each model has a card describing what it is best at, who built it, its context window, and what it costs to run. Models are labelled foundational (proprietary frontier models) or open source (open-weight models the network can ultimately serve itself).

Working across models

The interface is built around switching, not committing:

  • Switch models mid-conversation without losing context.
  • Compare two or more models on the same prompt, side by side.
  • Combine several models in one thread, so each answer can come from the model best suited to it.
  • Tune each request with a reasoning-effort control — spend more thinking on hard problems — and an optional web-search toggle.

Routing modes

You decide how much of the model choice to hand off:

Mode Who picks the model Best for
Custom You You already know which model you want
Auto The platform You want the right model chosen for you per request

Additional modes — optimising for reasoning depth, research, or credit efficiency — are on the roadmap.

Pricing

Model pricing is cost-plus-margin: the platform pays the underlying provider (or the network) for each request and adds a thin margin on top. There is no per-model markup game — value comes from usage volume, not from squeezing each call. See Tokenomics §4 for how that margin funds buyback-and-burn.

  • Text models are metered by tokens — the input you send plus the output you receive — at a per-model rate.
  • Image, video and other media models are metered per generation.
  • Every price is shown to you in stable, dollar-legible credits, whatever the model charges underneath.

Rates track real provider costs and are refreshed as those costs move, so the price you see reflects what a model actually costs to run rather than a fixed sticker.

How models get added

Today the catalog is curated: the team adds and maintains models, prioritising capability, coverage across labs, and genuine open-weight options. New flagship models are typically live within days of release.

The longer-term direction, set out in Tokenomics §5, is permissionless listing — anyone can publish a model, and protocol incentives flow to it by measured demand (credits burned, uptime, completion rate) rather than by popularity vote. Governance sets only the guardrails a model must meet; the market decides which models win.

Perspective AI does not try to be the cheapest way to call a model. It competes on breadth, privacy, uncensored access to open models, and freedom from being repriced or deplatformed.