Meituan released LongCat-2.5 on November 2, following September’s preview. The model combines visual understanding with multi-step tool use and targets complex software projects. Its official release notes list compatibility with coding-agent harnesses including Claude Code, Codex and Hermes.
Visual inputs meet sustained coding workflows
The practical pitch is a model that can interpret visual inputs while working on code, not simply produce isolated snippets. Meituan highlights interactive webpages, creative design and 3D tasks in its release announcement. These remain vendor-described capabilities rather than independently verified results.
API access and promotional pricing
Developer access is through LongCat’s API, with OpenAI- and Anthropic-compatible interfaces documented in its API overview.
Current promotional pay-as-you-go pricing is $0.30 per million uncached input tokens and $1.20 per million output tokens. The official pricing page describes these as limited-time discounted rates and says prices are subject to change.
Why it matters for builders
For small app teams such as SunMarc App Labs, LongCat-2.5 offers another model to evaluate inside existing coding workflows. A useful first test would provide a UI reference and one bounded app change, then compare the working result, test pass rate and total task cost.
Cheap tokens matter only if the agent completes the job reliably. A useful evaluation should judge the saved implementation, including whether it matches the reference and passes the project’s checks, rather than the model’s completion message alone.
For another approach to screenshot-driven workflows, see our coverage of visual and computer-using agents. The shared question is practical: can the model turn what it sees into a correct, reviewable change?