Meituan's LongCat 2.5 arrives on Vercel AI Gateway with vision and agent tool use

LongCat 2.5 lands on Vercel AI Gateway

Meituan's LongCat 2.5 is now available through Vercel AI Gateway, according to an announcement on the Vercel blog. The model pairs built-in image understanding with coding and multi-step tool use, and Vercel positions it for agents that work across large codebases and take on complete project builds rather than one-off edits.

LongCat 2.5 accepts both text and image inputs and offers a context window of one million tokens. Vercel says that capacity is meant to hold code, documents and the long task histories that accumulate when an agent works through many consecutive steps.

Where Meituan says the model shines

The capability claims come from Meituan itself. According to the Vercel post, the model's developer points to stronger performance in visual programming, creative design and 3D work, and says LongCat 2.5 can sustain agent workflows that stretch across the whole life of an engineering task — starting at the planning stage and running through to finished execution.

These are vendor claims rather than independent benchmark results, so developers will want to test the model against their own workloads before committing to it.

How to try it

The model is exposed under the identifier meituan/longcat-2.5, and Vercel offers two ways in:

• The AI Gateway model playground, for quick experiments with no setup required.
• Coding agents, including Claude Code, Codex and Cursor. Vercel's instructions are to install the latest Vercel CLI, run npx vercel ai-gateway setup, and then select meituan/longcat-2.5 inside the agent.

What the gateway layer provides

AI Gateway sits between applications and model providers, exposing a single API for calling models together with usage and cost tracking. Vercel lists automatic retries, provider failover and performance optimisations among its features, with the stated goal of higher availability than a single provider can offer on its own. The service also includes built-in custom reporting, zero-data-retention support, budgets tied to individual API keys and configurable routing rules.

Pricing is a notable part of the pitch. Vercel says AI Gateway matches the underlying provider's rates with no markup and no platform fee on inference, and that this policy extends to bring-your-own-key requests as well.

Why it matters

Model selection for coding agents has been dominated by a small group of providers, and a new entrant combining vision, tool use and a million-token context window gives developers a genuine alternative to evaluate. The vision capability in particular matters for agentic coding, where reading screenshots, design mockups and diagrams is often as important as reading source files.

The distribution route matters too. Because AI Gateway already fronts many models under one API and charges no markup on inference, the cost of trying LongCat 2.5 alongside existing favourites is close to zero: route traffic to it, benchmark the results, and route back without touching application code. For Meituan, appearing on a widely used developer platform is a way to put the LongCat family in front of an audience it might otherwise struggle to reach.

The usual caveats apply — the performance claims are the vendor's own, and real-world agent behaviour often diverges from model cards. But a multimodal, long-context model sitting one CLI command away is an easy experiment for most teams.