Meituan LongCat-2.5 officially released: coding capability significantly improved, benchmark score close to Opus 5

Beating AI News Flash: Meituan has released the official version of LongCat-2.5, more than a month after the preview version launched in September this year.

The model continues the previous generation's 1.6 trillion total parameters, approximately 48 billion activated parameters, and 1 million Token context, with a focus on improving visual programming and complex Agent task capabilities.

In official tests, the Terminal Bench v2.1 score improved from 70.8% in the previous generation to 88.4%, approaching Claude Opus 5's 89.1%. On the harder v4.0, it also rose from 3.0% to 38.5%, but still lags behind Opus 5's 53.9%.

The official version is now available for online experience and API access, and the model weights have not yet been open-sourced. Limited-time pricing is 2 yuan per million Tokens for uncached input and 8 yuan for output.

Source #Market Trends

Correction/Report

RelatedRecommended

Claude Opus 4.7 Review: Is It Worthy of the Title of Strongest Model?

Claude Opus 4.8 Discovers $4.5 Billion Bug, AI Era Witnessing Mass Production of Hackers

In the AI Security era, the capability will self-propagate, but the responsibility will not.

All Details About the Most Powerful Open Source Model DeepSeek V4: Performance on Par with Opus 4.6, Price Reduction, Encoding Benchmark Topping

V4-Pro-Max (Maximum Reasoning Power Mode) self-proclaims as the strongest open-source model to date, with its encoding benchmark reaching the top tier. The gap in inference and agent tasks with closed-source counterparts has significantly narrowed at the cutting edge.

Musk to Wage Price War, Grok Willing to Take Loss on $4.5 Million Deal with Claude - Is it Good News for SpaceX?

The biggest loss in the financial report comes from xAI, with the new model costing 25% less than Opus 4.8.

Claude Opus 4.8 Released, Anthropic Starts to Make "Trustworthiness" a Selling Point

Enhanced Self-checking and Multi-agent Orchestration for a More Trustworthy Model

On-Chain Activity