Meituan's LongCat 2.5 arrives in the Vercel AI Gateway with 1 million token context

Vercel added LongCat 2.5, Meituan's multimodal model, to the AI Gateway catalog. According to the company's changelog, the model combines native visual understanding with code generation and multi-step tool use, aimed at agents working on complex codebases and whole-project builds.

The technical highlight is the 1 million token context window, which accepts both text and image as input. In practice, this means fitting an entire medium-sized repository, extensive documentation, and the history of a long agent task into a single call, without needing to truncate or use RAG to maintain context.

What Meituan Promises in LongCat 2.5

Meituan points to specific improvements in visual programming, creative design, and 3D generation tasks, along with longer agent workflows that carry engineering work from planning through execution. There are no numerical benchmarks in the material published by Vercel comparing LongCat 2.5 with competitors like Gemini or Claude on coding tasks. This matters: a large context is not synonymous with better reasoning, just more room to put information into the input.

In short: what's documented is capacity (1M tokens, multimodal, agent-focused), not an output quality number. Anyone adopting the model in production will need to run their own evaluation on the specific use case before trusting it for critical tasks.

How to Call the Model via the AI SDK

The integration follows the same pattern as any AI Gateway model: just swap the model identifier. The name to use is meituan/longcat-2.5. A simple example with Vercel's AI SDK:

import { streamText } from 'ai';

const result = streamText({
 model: 'meituan/longcat-2.5',
 prompt: 'Outline a database migration in three brief steps.',
});

for await (const text of result.textStream) {
 process.stdout.write(text);
}

Anyone already using the AI Gateway to orchestrate calls to OpenAI, Anthropic, or other providers doesn't need to change anything in the architecture: just point model to the new identifier and test the behavior with your own project's prompts.

Setting It Up in Claude Code, Codex, and Cursor

To use LongCat 2.5 inside code agents like Claude Code, Codex, and Cursor, Vercel recommends installing the latest version of the CLI and running the gateway setup:

npx vercel ai-gateway setup

After setup, just select meituan/longcat-2.5 as the model inside the agent. It's the same flow used to switch between GPT, Claude, or Gemini in these editors: the AI Gateway acts as an intermediary layer, so the agent doesn't even need to know it's talking to a Meituan model.

Why Go Through the AI Gateway Instead of the Provider Directly

Vercel's AI Gateway exists to solve a common problem for those building with multiple models: managing keys, cost, and failover scattered across different providers. It offers:

• a unified API to call any model in the catalog, including LongCat 2.5
• usage and cost tracking per API key, with configurable budgets
• automatic retries and failover for higher uptime than a single provider alone
• routing rules between models and support for Zero Data Retention
• pricing equal to the original provider's, with no markup and no platform fee on inference, including requests made with your own key (BYOK)

This last point is what sets the AI Gateway apart from a simple proxy: Vercel doesn't charge a toll on processed tokens, so the cost of running LongCat 2.5 is exactly what Meituan charges at the source.

What Remains Open

The changelog doesn't include a per-token pricing table or a latency comparison with other long-context models, so anyone deciding to put LongCat 2.5 into production will need to measure this in their own project, both in terms of cost and response time for prompts that use a significant portion of the 1 million tokens available. It's worth testing first in Vercel's model playground before swapping the model in an agent workflow that's already in production, since large context tends to increase latency and cost even when the per-token price is low.

For those already running code agents with smaller context windows, the practical question is whether the task really needs 1M tokens or whether a cheaper model, with well-built RAG, solves the same problem with less latency. LongCat 2.5 expands the options within the AI Gateway, but choosing the right model remains a matter of measurement, not assuming that more context is always better.