LongCat 2.5 Preview Makes Million-Token Agent Tests Cheap
Meituan’s new multimodal model arrives with one-million-token context, OpenAI and Anthropic API compatibility, and pricing low enough to justify evaluation—but not a production migration.
What launched
Meituan made LongCat 2.5 Preview available through its API on September 25, 2026. Its documentation lists image understanding, code generation and analysis, automated programming, and integrations with Claude Code, OpenCode, Hermes, OpenClaw and Kilo Code.
The API supports OpenAI and Anthropic formats, a one-million-token context window and up to 128,000 output tokens. A first evaluation therefore does not require a new integration architecture.
Why operators care
Vercel AI Gateway currently lists the model at $0.30 per million input tokens and $1.20 per million output tokens. That is low enough to justify real internal workloads and shadow traffic rather than another isolated demo.
Long context is not automatically economical, however. Unnecessary tokens, retries, failed tool calls and tail latency can erase the headline saving quickly.
Our deployment call
We would add LongCat 2.5 Preview as an experimental route behind an existing model gateway, with strict context limits and no authority over critical actions. It should first face the incumbent on the same repositories, documents and tools.
Production traffic follows only if it lowers cost per completed task without reducing reliability. API compatibility makes the trial easy; the Preview label means migration should wait for evidence.
LongCat 2.5 Preview matters less because it is another large model and more because it makes a million-token, multimodal agent API cheap enough for serious testing and easy to attach to an existing stack. Vercel currently lists it at $0.30 per million input tokens and $1.20 per million output tokens, while Meituan exposes OpenAI- and Anthropic-compatible formats. That is enough for Virtual Arc to test it now, but not enough to make it a production default. We would put it behind our model router, cap context aggressively and run shadow traffic on bounded coding and document workflows against the incumbent. We would score cost per completed task, retry rate, tool-call accuracy and p95 latency—not token price or vendor claims. If it wins those evaluations across two release cycles, we would route low-risk volume to it and use the result to renegotiate existing spend. Until then, the Preview label and limited independent evidence make a full migration an avoidable reliability bet.