LongCat 2.5 brings visual work into its coding API

A screenshot can communicate a layout problem that takes several paragraphs to describe. LongCat’s November 2 API release combines visual understanding with coding and tool-based agent work, according to its platform changelog. The same record dates LongCat-2.5-Preview to September 25; the newly listed release is a separate event.

For someone building a small website, the potential use is to pair a visual reference with the files that implement it, then review a proposed change. The vendor lists compatibility with agent tools including Claude Code, Codex and Hermes. That describes an integration option, not evidence that a full project will work without supervision.

A large context still needs a bounded task

The parameter reference lists a 1,048,576-token context, counting input and output together, and a separate 262,144-token maximum output. Those are ceilings, not a recommendation to fill a request with every file in a repository.

For a proposed evaluation, copy a small page and provide its screenshot plus the relevant layout files. Ask for one concrete repair, such as keeping a form inside the phone viewport. Compare the resulting page at both sizes and inspect the changed files for unrelated edits. This is an unexecuted exercise; CreatorStack did not generate code or score the model.

Budget the work across turns

The checked pricing page lists temporary discounted rates per million tokens: $0.30 uncached input, $0.006 cached input and $1.20 output. It names platform prices and billing records as authoritative and gives no guaranteed discount end date.

As illustrative arithmetic, usage totalling 100,000 uncached input tokens and 10,000 output tokens would cost $0.03 + $0.012 = $0.042 at those rates. That excludes additional agent turns, tool services and other account charges; it is not a measured price for repairing a website. Repeatedly sending context can matter more than the first request’s price.

The API overview describes LongCat’s own authenticated service with OpenAI- and Anthropic-compatible formats. A compatible format does not make this an OpenAI or Anthropic model, or guarantee identical behavior in every client. The account FAQ distinguishes mainland and international access instructions; Malaysia account eligibility and payment were not tested.

Use the AI output evaluation guide to define the result you need before switching a working setup. Begin with one reversible task, then compare the change, usage record and reader-facing behavior.