Meituan Releases LongCat-2.5, a Multimodal Model Built for Long Coding Jobs
Meituan has made LongCat-2.5 generally available on its developer platform, turning a September preview into a full release aimed at long, multi-step coding and agent work. The model arrives with native multimodality, a focus on whole-project engineering tasks, and pricing that undercuts most frontier models from US labs. It also arrives with important caveats: its headline benchmark claims have not yet been independently verified, and its weights have not yet been released for self-hosting.
From preview to general release
LongCat-2.5 first appeared as a preview on September 25, 2026. On November 2, 2026, Meituan listed LongCat-2.5 as available on the LongCat API platform, according to the platform changelog. The changelog describes three core features for the general release: native multimodality that combines visual understanding with agentic coding for visual coding, creative design, and 3D tasks; long-horizon agents for complex engineering, including work on large codebases and full-project builds; and compatibility with agent harnesses such as Claude Code, Codex, and Hermes.
Meituan is best known as a Chinese food delivery and local-services company. LongCat-2.0 was released in June 2026, and LongCat-2.5 extends that line with a stronger emphasis on sustained, tool-driven work rather than single-question answers: an agent that can plan, write code, run it, inspect the result, recover from failures, and keep going until the job is done.
Scale, context, and price
LongCat-2.5 is described in preview and launch coverage as a mixture-of-experts model with 1.6 trillion total parameters and about 48 billion parameters active for each token. That distinction matters. In a mixture-of-experts design, only a small subset of the model is used for any given token, so the active figure, not the total, is the closer guide to inference cost. Reports also describe a 1 million-token context window and a maximum output of 128,000 tokens, figures that trace back to the preview announcement. A context window of that size lets an agent hold a large codebase, long documents, or a multi-stage workflow in working memory at once.
On price, the LongCat pricing page lists LongCat-2.5 as the only model on its pay-as-you-go plan, at $0.30 per million uncached input tokens, $0.006 per million cached input tokens, and $1.20 per million output tokens. Those rates are marked as a limited-time discount, and the page gives no end date for the discount. It also notes that prices may change and that the platform billing records are the final word. For teams that run coding agents at volume, that combination of a large context window and low per-token pricing is the main reason to test the model, especially on routine work such as refactors, test writing, and documentation where token costs add up quickly.
What Meituan says it can do
Meituan reports that LongCat-2.5 scores 88.4 on Terminal Bench v2.1, compared with 70.8 for LongCat-2.0, according to coverage of the company benchmark table. Separate coverage reports that Meituan positions LongCat-2.5 between Claude Opus 4.8 and Claude Opus 5 across several mainstream evaluations, surpassing Claude Opus 5 on select tasks. Those are company-reported results from in-house testing, and as of November 2, 2026, no independent evaluator had published a third-party assessment of LongCat-2.5. Readers should treat the benchmark claims as directional until independent testing appears.
One training detail stands out. Meituan says it trained the model on extended sequences of agent actions, including failed attempts turned into training tasks, so the model learns from real failure modes rather than only from curated examples. Coverage also reports that the model was trained on domestically produced Chinese AI accelerators, continuing the approach used for LongCat-2.0. That detail matters for the wider industry story: it shows a major Chinese lab building a large agent model without relying on the usual US chip supply chain.
How to access it, and what is still missing
LongCat-2.5 is available through Meituan online chat service and its API platform. Launch coverage notes that GitHub and Hugging Face releases are planned, but as of November 3, 2026, no LongCat-2.5 repository or licence had appeared on those platforms, so the model cannot be self-hosted today. That limits who can use it in production. Teams that need to run a model inside their own infrastructure, or that need detailed data handling guarantees, will have to wait or route through a gateway.
Why it matters
LongCat-2.5 matters less as a single model and more as a sign of where agent pricing and competition are heading. A model built for long-horizon coding work, with a very large context window and a low promotional price, puts pressure on what teams pay for AI tooling even if they never use this model directly. It also shows that credible agent models are now coming from outside the usual US labs, with different hardware, different pricing, and different trade-offs.
The sensible response is testing, not switching on headlines. Run two or three routine tasks on LongCat-2.5 alongside your current model in a test repository with no sensitive code, and compare results, review time, and token cost.
Sources
- LongCat API Platform changelog: LongCat-2.5 Now Available, November 2, 2026
- TechTimes: Meituan Releases LongCat 2.5 on Chinese Chips; Benchmark Claims Await Independent Tests
- The DevDrift: Meituan LongCat-2.5, low-cost model for coding agents
- AI Affairs: Meituan releases LongCat-2.5 for long-running coding tasks
