Meituan releases LongCat-2.5 for visual coding with a 1M-token context

Meituan made LongCat-2.5 available as a full release on November 2, positioning the model for visual coding, design and 3D tasks. Its API documentation now identifies LongCat-2.5 in request examples and publishes separate rates for input and output tokens. The release combines visual understanding with coding agents that carry out extended engineering tasks.

It follows LongCat-2.5-Preview, released on September 25 with image understanding and development-tool integrations already included. The November 2 changelog emphasizes complex codebases and full-project builds. Developers can access the model through existing coding tools configured to call LongCat.

The company’s technical blog describes 1.6 trillion total parameters, with approximately 48 billion activated per token. Meituan says it turned real agent failures into reproducible training tasks.

Meituan reports a Terminal Bench v2.1 score of 88.4 for LongCat-2.5, compared with 70.8 for LongCat-2.0. The stated setup uses Claude Code 2.1.280, a maximum sequence length of 512K and an average over three runs. These are developer-reported results, which AI Ping has not reproduced. Scores from other benchmark versions in the table cover different evaluations.

A shared budget for input and output

The model specification lists a context length of 1,048,576 tokens, covering input and output combined. The maximum output setting is 262,144 tokens. Large code and image inputs therefore share the request’s capacity with the generated response.

Supported images include PNG, JPG, JPEG, WebP and GIF. They can be supplied through a public URL or inline Base64 data, with a maximum of 10 MB per image and 50 images per request. SVG and HEIC/HEIF are excluded. The same limits apply when combining screenshots and photographs in a request.

Videos must be supplied by URL. Each file may be up to 100 MB and must be shorter than 3,600 seconds, with up to ten videos per request. The transport requirements therefore differ between inline image inputs and video inputs.

LongCat documents an Anthropic-compatible interface for Claude Code and a Responses interface for Codex. Both require a LongCat API key and quota, with examples selecting LongCat-2.5. The Claude Code guide sets an output allowance of 131,072 tokens: a client configuration example distinct from the model’s maximum. The configured service address and model determine the response provider.

The current limited-time rates per million tokens are $0.30 for uncached input, $0.006 for cached input and $1.20 for output. These categories are billed separately within a request. The price page gives no promotion end date. It does not announce a reduction from the preview’s rates.

Pay-as-you-go uses a prepaid balance and requires account verification before topping up. Successful HTTP 200 responses incur token charges. Paid balances do not expire, and promotional credits are consumed first. Listed payment methods include credit and debit cards, Google Pay, Alipay and WeChat Pay. The platform’s displayed prices and billing records take precedence.

Token Packs are a separate product: a purchased quota valid for 30 days. They are accounted for independently of the prepaid balance. When an account holds both, its Token Pack quota is consumed first. For Token Packs, cache-hit inputs do not consume the quota; cache-miss inputs and generated outputs do.