Vercel's AI Gateway offers Cat Jam reasoning model free during stealth period

Vercel has added a reasoning model called Cat Jam to its AI Gateway, and it costs nothing for as long as the model remains in stealth, according to a changelog post the company published on October 15. The model targets programming work and agent-driven workflows, with inference tuned for fast responses.

What the model does

Cat Jam is a reasoning model: responses include the model's intermediate reasoning text alongside the final answer. Developers can select from five reasoning effort levels — low, medium, high, xhigh and max — which offers a knob for trading thoroughness against speed.

The model supports function calling, so agents built on it can be wired to a developer's own tools, and it produces structured outputs that conform to a JSON schema supplied with the request. There are notable constraints. Cat Jam takes text input only, ruling out images or other modalities, and Zero Data Retention is not offered for this model. Vercel states that prompts may be kept for safety reasons, which effectively rules it out for workloads with strict confidentiality requirements while it is in stealth.

How to access it

The model identifier is stealth/cat-jam. According to Vercel, it works with the AI SDK, the OpenAI-compatible Chat Completions and Responses APIs, the Anthropic-compatible Messages API, and coding agents connected to AI Gateway. That means tools such as Claude Code, Codex and Cursor can run it: install the latest Vercel CLI, run npx vercel ai-gateway setup, then pick stealth/cat-jam from the agent's model selection. Vercel also exposes the model in its playground for quick experiments without setting up an agent.

The gateway context

AI Gateway is Vercel's unified entry point for calling many language models from one API. It handles usage and spend tracking, retries, failover, and performance tuning that Vercel says keeps availability higher than connecting to a provider directly. Custom reporting, per-API-key budgets and routing rules are also built in.

On pricing, Vercel passes provider rates through without adding a markup or a platform fee on inference, including for requests made under a bring-your-own-key arrangement. That framing positions the gateway as neutral infrastructure rather than a model vendor in its own right.

Why it matters

A free window on a reasoning model is a low-risk way for developers to evaluate it against what they already run, especially in agentic coding, where reasoning-capable models are popular but token costs accumulate quickly. The combination of function calling, JSON-schema outputs and low latency covers the core requirements of most agent frameworks, so integration effort is minimal.

The retention caveat deserves as much attention as the price. Because prompts may be stored, teams should not route proprietary code or sensitive data through Cat Jam while it is in stealth, and should treat the free period as evaluation time rather than production-ready capacity.

The release also shows how model distribution is shifting. Vercel does not name the company behind Cat Jam, consistent with its stealth label, and instead bets that developers will standardize on a single gateway API and swap models in and out as offers like this appear. Free stealth access is becoming a cheap customer-acquisition channel for new model providers, and gateways that aggregate them are the place those experiments happen.