Meituan's LongCat 2.5 Lands on AI Gateway: 1M Context + Visual Coding

Title

Meituan's LongCat 2.5 Lands on AI Gateway: 1M Context + Visual Coding

Meta Description

Meituan's LongCat 2.5 launches on AI Gateway, combining native visual understanding, coding, and multi-step tool calling, with text and image input and a 1M-token context window. It can be invoked via the model name meituan/longcat-2.5 in tools like AI SDK, Claude Code, and Cursor.

Article Content

Meituan's large model LongCat 2.5 has officially landed on AI Gateway. The model integrates native visual understanding, coding capability, and multi-step tool calling, aimed at agents working in complex codebases and full-project builds.

A Model Built for Agentic Engineering Workflows

LongCat 2.5 is not positioned as a mere code completion tool. According to Meituan's description, the model shows clear improvements in visual coding, creative design, and 3D tasks, while supporting longer agentic workflows—pushing engineering work from planning all the way through to execution. This means it is better suited to continuous tasks like "breaking down requirements through to implementation" rather than just answering isolated questions.

The model supports both text and image input, with a context window of 1 million tokens, covering code, documents, and longer task histories. For agent scenarios that need to repeatedly reference project files and track dependencies across files, long context is a hard requirement—this specification directly determines whether it can maintain coherent reasoning across an entire codebase.

Access: Invoke with a Single Model Name

In AI Gateway, invoking the model simply requires the model name meituan/longcat-2.5. The official AI SDK example shows the basic pattern for streaming calls: import streamText, specify the model name and prompt, then read the text stream chunk by chunk. For teams with existing AI SDK workflows, this integration approach has a very low migration cost.

To use it in coding agents like Claude Code, Codex, and Cursor, first install the latest version of Vercel CLI and run npx vercel ai-gateway setup, then select meituan/longcat-2.5 in the agent. The official playground is also available for direct trials.

AI SDK and streamText

AI SDK is an open-source TypeScript library from Vercel for unified access to large models in applications, abstracting away differences between different vendors' APIs. streamText is one of its core functions, issuing text generation requests in streaming mode: as the model returns content, it is pushed to the client chunk by chunk rather than waiting for the entire generation to complete before returning at once. Streaming output is especially important for coding agents because generating code can take tens of seconds, and chunk-by-chunk rendering lets users see progress during generation and interrupt at any time. The object returned by streamText includes an iterable text stream and also supports aggregating the complete result and counting token usage after the stream ends, making it easy to integrate usage and cost tracking.

BYOK (Bring Your Own Key)

BYOK stands for Bring Your Own Key, meaning users use the API keys they obtained from model vendors and forward requests through the gateway, rather than using keys uniformly provided by the gateway platform. For teams, the benefit of BYOK is that billing occurs directly between the user and the vendor, making it easy to continue using existing contract prices and quotas; the gateway only handles routing, retries, and usage observability. AI Gateway also does not mark up or charge platform fees for BYOK requests, meaning teams can use the gateway as a pure access layer without worrying about additional cost stacking.

AI Gateway's Role: Unified Entry Point with Controllable Costs

AI Gateway itself provides a unified API for invoking models, tracking usage and costs, and configuring retries, failover, and performance optimization to achieve higher availability than a single vendor. It also has built-in capabilities such as custom reports, zero data retention support, API key budgets, and routing rules.

On billing, AI Gateway reflects vendor pricing as-is, with no markup and no platform fees for inference, and the same applies to Bring Your Own Key (BYOK) requests. For teams that need to integrate multiple models while wanting unified cost and routing management, the value of this gateway layer lies in decoupling "choosing a model" from "managing models."

Failover and Routing Rules

Failover means that when the primary model request fails or times out, the gateway automatically reroutes the request to a preconfigured backup model, avoiding business interruption caused by a single point of failure. Routing rules allow decisions about which model to use based on request characteristics (such as task type, cost ceiling, user group). Combined, these let teams hard-code a single logical model name in their code while the gateway layer decides which vendor's model is actually used behind the scenes. This is especially useful for multi-model comparisons, gradual rollouts, and vendor rate limiting, and it also turns "choosing a model" from a code change into a configuration change.

A Few Directions Worth Watching

Visual coding and 3D tasks were specifically called out by Meituan, indicating targeted optimization between multimodal input and code generation—practically meaningful for workflows involving UI reproduction and design-to-code. Multi-step tool calling combined with long context points toward a more autonomous agent form—the model needs to decide which tools to call, how to chain steps, and how to keep its goal from drifting over long tasks.

However, the information officially disclosed so far focuses on capability directions and access methods, lacking concrete benchmark data. The actual performance of visual coding and 3D tasks, as well as effective recall quality under a million-token context, still need validation in real projects. For teams planning to replace or add models in coding agents, starting with small-scale comparative testing in the playground is the more prudent path.

Effective Recall Under a Million-Token Context

Context window refers to the maximum number of tokens a model can process in a single request; 1 million tokens is roughly equivalent to thousands of pages of text or a medium-sized codebase. But window capacity and effective utilization are two different things: as input grows longer, the model's attention to information in the middle tends to decay, a phenomenon known in the industry as "lost in the middle." Therefore, when evaluating long-context models, beyond the window specification, it is also important to look at actual recall accuracy on tasks such as long-document question answering and cross-file dependency tracking, as well as inference latency and cost under long inputs. These metrics have not yet been disclosed officially.

One-line Summary

Meituan's LongCat 2.5 lands on AI Gateway, focused on visual understanding, coding, and multi-step tool calling.

Paragraph Summary

Meituan's large model LongCat 2.5 is now officially available on AI Gateway, positioned as a model for agentic engineering workflows that integrates native visual understanding, coding capability, and multi-step tool calling. The model supports text and image input with a 1M-token context window, covering codebases, documents, and long task histories. Invocation only requires the model name meituan/longcat-2.5, and it can be configured for use in coding agents like Claude Code, Codex, and Cursor. AI Gateway provides a unified API, usage tracking, retries and failover, budgets, and routing rules, with billing reflecting vendor pricing as-is and no markup. Official benchmark data has not yet been published, and the effects of visual coding and long context still need real-world validation.