- Vercel says Kimi K3.1 Pro and a faster Highspeed version are on AI Gateway, selected by changing a model name.
- When a model is a name in a setting, choices about cost and data handling move to whoever edits that setting.
- Ask who can change model names, whether budgets are set per key, and what happens when promotional rates end.
Choosing an AI model is starting to look like changing a setting. This week's notices from Vercel show why that deserves attention from the people who sign the budgets. When a model is just a name in a configuration file, the decision about which model touches company data can move to whoever edits that file.
What Vercel announced
Vercel's gateway now lists two Kimi models from Moonshot AI: K3.1 Pro, and a Highspeed variant built for interactive use where waiting is a cost. K3.1 Pro is the opening release of Moonshot's K3.1 line.
Vercel's notice lists gains over the earlier Kimi K3 in four areas: writing code, 3D modeling, financial analysis and agent work. Agent work means software handling a long chain of steps without a person guiding each one. The model keeps a 1M-token context window, which is how much text it can weigh at once, and it accepts images. The notice also says the model gets facts wrong less often than older Kimi versions when prompts are hard.
The notice gives no test results or price for the model. These are Vercel's description of the release, not independent findings. Moonshot AI intends to publish the model weights, the trained files that define it, by November 4. Vercel adds that support from more providers will follow.
How the switch works
AI Gateway is a single service that sits between a company's software and many AI providers. Vercel describes one interface for calling models, watching usage and spend, and setting how requests are retried, redirected or routed. Failover means sending a request elsewhere when a provider is down.
To use the new model, a developer sets the model name to moonshotai/kimi-k3.1-pro. The faster version is moonshotai/kimi-k3.1-pro-highspeed. Developers can reach either one from Vercel's AI SDK, from three API styles (Chat Completions, Responses and Anthropic Messages), or from a coding agent linked to the gateway. Agents such as Claude Code, Codex and Cursor can select the same names after a Vercel CLI setup.
That is the mechanism worth noticing. The software around the model stays the same. Only a string of text changes.
Six models in four notices
Vercel published four changelog notices between October 19 and October 23. Together they cover six models. That is one publisher's feed over a few days, not a measured trend. The details still show what a switch involves.
Three of the four notices carry promotional terms. The Kimi notice does not. MiniMax M3.1 Flash was free on the gateway until October 22. The notice says that price held only if MiniMax itself handled the call. To guarantee that, developers had to set the provider order. Once the offer ended, Vercel says, pinned requests bill at MiniMax's standard rate.
Mercury 3, from Inception, launched at 50% off, according to Vercel. Inception's own figure for its output is more than 1,000 tokens per second. Baseten's Mammoth-1 and Elephant-1 are free through October 27. Vercel describes them as decision models that sort text, route requests and grade content against criteria you set.
Where the risk sits
The lesson here is that the model has become a setting, not a contract. A contract has an owner, a review and a renewal date. A setting has whoever last edited it.
Consider the free offer. A request pinned to one provider to capture a free rate can keep that pin after the rate changes. The code works the same way before and after. Only the invoice differs.
Cost is the visible part. The Kimi notice states that the gateway passes through provider prices with no markup and no platform fee on inference, bring-your-own-key traffic included. So the bill follows each provider's rate card, and someone has to read it.
Data handling is the quieter part. Vercel lists Zero Data Retention support among the gateway's features. The Kimi notice does not say which models or providers that covers. A model that is cheap or free is not useful if it sends data somewhere your policy forbids.
Decision models add another judgment. Per Vercel, they answer with a choice, a probability and a score. The application then decides how high a probability must be before it acts alone and when a person reviews the result. That threshold is a business rule hidden in a number.
Questions to put to your team
Who can change a model name in production, and who can change it inside a coding agent? Are budgets set per API key, which Vercel says the gateway supports? Which requests are pinned to a provider for a promotional rate, and what happens when it ends?
Does Zero Data Retention apply to the exact model and provider in use? Who owns the threshold that decides when a machine acts without review? Has anyone tested a new model on your own tasks, rather than relying on a vendor's description?
A model name is easy to change. The cost, the data path and the accountability are not, and each of them needs a named owner.
Produced by the WebPulse Newsroom with AI assistance from the original reporting credited below, and checked against that source by our editorial review. How we use AI.
Original reporting: Vercel.
