LongCat-2.5-Preview became available on Meituan’s LongCat API Platform on September 25, according to the official LongCat API Platform change log. The date matters because TechNode published its account on September 30, five days after the platform’s version entry. The release record centers on an API offering, image understanding, and coding-oriented functions. Neither the change log nor TechNode’s account supplies independent benchmark results, a parameter count, context length, latency data, or a description of the underlying architecture. Those omissions set the boundaries for assessing the preview release.
Image Understanding Enters the LongCat-2.5-Preview API
The change log presents image understanding as the new capability in LongCat-2.5-Preview. It says the model can parse image content and support cross-modal question answering, content summarization, and complex visual reasoning. These are platform documentation claims from LongCat, not results from a disclosed testing process. TechNode similarly describes answers about images, visual-content summaries, and more complex visual reasoning as functions reported in the documentation. The available material establishes the documented feature set but does not quantify accuracy, compare the model with another system, or specify a test dataset.
The scope of the entry is also deliberately narrow. LongCat’s documentation identifies image understanding, while the materials reviewed for this release offer no verified modality matrix and no claim covering video or audio. For developers deciding whether an image input belongs in a workflow, that leaves the API documentation as the operative guide. It also keeps the release distinct from a broad assertion about every form of multimodal input. EastFrontier previously covered a vision feature through an API platform in its report on DeepSeek V4-Flash; LongCat’s September entry has to be read on its own documented terms.
The release is labeled Preview, and the available documentation does not describe a public launch event, geographic rollout, adoption totals, or revenue contribution. It does establish a versioned API-platform availability date and a set of stated image-related functions. That is a useful operational distinction for anyone tracking model releases through product documentation rather than promotional claims or inferred market uptake.
Coding Functions and Listed Developer Environments
Alongside image understanding, the official change log characterizes LongCat-2.5-Preview as having stronger coding capability. It lists code generation, code understanding, and automated programming tasks among the relevant functions. The language is Meituan and LongCat’s own description of the model, so it does not amount to an independent finding that the preview outperforms a named rival or meets a particular benchmark. TechNode reports improved coding performance in the same documentation-based frame.
The change log says the model is designed to work with established development environments. It gives Claude Code, Hermes, OpenClaw, OpenCode, and Kilo Code as examples. Compatibility should be read as a documented connection path, not ownership by Meituan of those separately named tools. It also does not establish that every possible workflow within each environment has been tested. The meaningful detail is that the release documentation puts a coding model and developer environments in the same API product description.
A broader official tools overview lists integrations across coding agents and desktop clients, including Codex, Claude Code, Cline, and other developer-facing tools. The page says the platform’s metered API plan and Token Pack plan can be used with supported AI coding tools. Its generic connection rule covers tools that support the OpenAI or Anthropic protocol and permit a custom provider. Users create an API key on the LongCat Platform, then can use the documented OpenAI-compatible endpoint at https://api.longcat.chat/openai/v1 or the Anthropic-compatible endpoint at https://api.longcat.chat/anthropic.
API Billing Defines the Immediate Commercial Terms
LongCat’s official pricing page lists LongCat-2.5-Preview among models covered by its metered billing. It gives limited-time discounted rates per 1 million tokens of $0.30 for uncached input, $0.006 for cached input, and $1.20 for output. The same page lists ¥2 for uncached input, ¥0.04 for cached input, and ¥8 for output. The platform warns that the figures may change and identifies its platform and billing records as authoritative, so the listed rates should be treated as current documentation rather than a lasting price commitment.
The tools overview names LongCat-2.5-Preview and LongCat-2.0 as the models it currently supports. That pairing links the preview to an earlier model line that EastFrontier examined when Meituan unveiled LongCat-2.0. The current documentation, however, supports a more limited conclusion about 2.5-Preview: API-platform access, an API-key connection process, documented developer-tool pathways, and stated image and coding capabilities. It does not establish the licensing terms of this particular preview release.
For a developer, the immediate questions have concrete documented answers: LongCat-2.5-Preview is available through the API platform; a platform API key is required for compatible-tool connections; the documentation offers OpenAI- and Anthropic-protocol routes; and the pricing page specifies discounted token rates subject to change. The same documents leave performance validation, actual tool behavior in individual setups, and commercial adoption outside the record. Meituan’s release therefore provides an identifiable API product update with a declared multimodal and coding focus, while its performance claims remain the company’s own documentation claims.