Perplexity's Decider V1.2 models arrive on Vercel's AI Gateway

Perplexity's Decider models arrive on Vercel's AI Gateway

Vercel has added two of Perplexity's decision-oriented models to AI Gateway, its service for calling hosted AI models with a single key. According to the Vercel changelog, Decider V1.2 27B and the smaller Decider Flash V1.2 4B are now available through the gateway.

The Decider family is not a general chat model. Vercel describes the pair as multimodal decision models built to classify content, steer requests to the right destination, and score inputs that can be text, structured records or images.

Typed questions, probabilistic answers

The core mechanic is that the models take typed questions and return probabilities. The changelog breaks the question types into three:

• Boolean questions return the probability of a yes answer.
• Choice questions pick from a set of categories the developer defines up front.
• Score questions grade content against an ordered rubric.

Because answers are typed and numeric, an application can act on them directly — steering a support ticket to the right queue, flagging a damaged product from a photo, or pushing a low-confidence call to a human reviewer. Several questions can also be evaluated against the same input within one request, so an app that needs multiple judgements about a single document does not pay for repeated round trips.

Images are supported too: developers embed base64 image_url objects in JSON state through the AI SDK or the TypeSafe-compatible API, with Perplexity's own documentation covering request formats and question types.

Two sizes for two kinds of work

Decider V1.2 27B, exposed as perplexity/pplx-decider-v1.2, is the larger of the two. Vercel positions it for fine-grained calls and elaborate scoring rubrics — situations where the evaluation itself is subtle.

Decider Flash V1.2 4B (perplexity/pplx-decider-flash-v1.2) is the compact sibling, aimed at bulk labeling, filtering and routing where the criteria are straightforward and volume is high. That is the profile of work that tends to fire on every request or every piece of user-generated content, where model size and latency add up quickly.

Both models can be tried in Vercel's model playground, and the changelog walks through the same delivery-report check three ways: through the AI SDK, Perplexity's Decisions API, and the TypeSafe-compatible API.

Gateway integration and billing

AI Gateway's pitch is consolidation — one key for the whole catalog, with usage and spending visible in a single view. Vercel notes that decision requests land in the gateway's request logs and count toward spend budgets, so routing and classification traffic shows up in the same observability pipeline as the rest of a project's model calls.

On cost, the company says inference is billed at the provider's own rates with nothing added on top, and that the same terms apply when customers plug in their own provider credentials via BYOK.

Why it matters

A large share of how language models end up in production is quietly decisional: should this ticket escalate, is this image acceptable, which model should handle this query. Today that work is typically done by prompting a general-purpose chat model and parsing free text back out of it — workable, but brittle. Purpose-built decision models with typed questions and probability outputs offer a cleaner contract: thresholds map directly onto routing logic, and there is no ad-hoc parsing layer to break.

For teams already routed through Vercel's gateway, the addition means classification and routing decisions can live inside the same keyed, logged and budgeted pipeline as the rest of their model traffic. And with provider pricing passed through unchanged, the value here is operational rather than a margin play.

The two-size split also matches how these workloads behave in practice: heavyweight judgement where nuance pays for itself, and cheap fast filtering everywhere else. Developers now get both ends of that spectrum behind one interface — assuming the models deliver on that positioning.