A food-delivery giant has shipped a trillion-parameter agent model with budget-API pricing. The numbers are impressive, but most of them come from Meituan itself.
Meituan LongCat-2.5 is the latest sign that frontier-scale AI is no longer the preserve of dedicated labs. The Chinese super-app, best known for food delivery, released the model on 2 November, pitching it at long-running coding agents rather than one-shot chat. For Indian developers the headline is the price: $0.30 per million input tokens. The catch is that most performance claims are Meituan’s own.
What Meituan LongCat-2.5 actually is
LongCat-2.5 is a mixture-of-experts model with 1.6 trillion total parameters, of which about 48 billion activate for each token. Meituan calls it natively multimodal: text and images are handled in a single token format instead of through a bolted-on vision module. It is built for “long-horizon” work, meaning agents that must keep a terminal, browser or spreadsheet in mind across many steps.
Model listings show a context window of about one million tokens and up to 131,072 tokens of output. It succeeds LongCat-2.0, which Meituan released earlier this year with open weights under an MIT licence. Meituan also says a 4-bit inference setup lifts prefill throughput by 43.3% and decode throughput by 42.77% over an 8-bit baseline.
Benchmarks: strong, but not a clean win
The table below uses scores from Meituan’s own launch post.
LongCat-2.5 vs rivals: coding-agent benchmarks
Scores as reported by Meituan. Bars compare each score with the best in that row.
| Benchmark | LongCat-2.5 | LongCat-2.0 | Claude Opus 5 | GPT-5.6-Sol |
|---|---|---|---|---|
| Terminal Bench v2.1 | 88.4 | 70.8 | 89.1 | 88.8 |
| Terminal Bench v4.0 | 38.5 | 3.0 | 53.9 | 37.3 |
| DeepSWE v1.1 | 64.6 | 9.7 | 68.8 | 72.7 |
★ = best score in the row. Source: Meituan LongCat-2.5 launch post (vendor-reported, not independently verified). Higher is better.
The jump over LongCat-2.0 is huge. Against rivals the picture is mixed: on Terminal Bench v2.1 it sits within a point of the best scores, but on the harder v4.0 test it trails Claude Opus 5 by a wide margin. It does beat Claude Opus 4.8 on DeepSWE (64.6 versus 59.0). We could not find independent evaluations yet, so treat every number here as a vendor claim.
Pricing and access in India
LongCat’s API lists $0.30 per million input tokens, $0.006 for cached input and $1.20 per million output tokens. Trackers describe this as a limited-time launch discount, so the price may rise. It matches the list price OpenRouter shows for DeepSeek V4.1 Flash, and Meituan’s own table credits that model with 90.6 on Terminal Bench v2.1, ahead of LongCat-2.5.
TechNode reports that GitHub and Hugging Face releases are planned, but we found no LongCat-2.5 repository yet. Even when weights arrive, self-hosting is unrealistic for most Indian startups: the LongCat-2.0 model card recommends 16 H20 GPUs.
What to check before you build on it
Data handlingPrompts go to a Chinese provider. Read the privacy terms and check what your clients’ compliance rules allow.
Price after the discountConfirm the standard rate before committing a product to it.
Payments and supportCheck whether the platform accepts payment from India.
Your own codeRun it on a real repository. Benchmarks rarely match your stack.
Bottom line: LongCat-2.5 looks like a serious, cheap option for coding agents, and worth a weekend test. Wait for independent benchmarks and the open weights before betting on it.
⚠️ Disclaimer: Benchmark scores are reported by Meituan and have not been independently verified. Pricing and availability may change, so check the official LongCat site before you build or buy. This article is for information only.
📚 Sources: