LongCat-2.5: Multimodal Agents, API Access and Pricing

LongCat-2.5 is Meituan's native multimodal model for long-horizon agent tasks, focused on coding from images, complex engineering and sustained tool use. It goes beyond a single text response: in an environment with a browser, terminal and testing tools, it can repeatedly modify, run and inspect results toward the same objective. Official introduction

Guishu Intelligence's live catalog currently lists the exact model ID LongCat-2.5 with text and image input support. A catalog listing does not establish that every client, tool or long-running task has been tested. Before integrating, consult the model guide and live pricing. Information in this article was checked on 2026-11-03.

Specifications and version: the release is no longer called Preview

Item	Current information
Model ID	LongCat-2.5; preserve capitalization and the ASCII hyphen
Model size	1.6 trillion total parameters, with about 48 billion activated per token
Official context limit	1,048,576 tokens, or 1M, counting input and output together
Official maximum output	max_tokens limit of 262,144, or 256K
Main uses	Visual coding, interactive 3D, long-horizon coding and tool tasks
Release dates	The official announcement is dated November 1; the API change log records availability on November 2

Sources: official introduction, model parameters, change log. The current parameter documentation specifies 256K maximum output, not the 128K mentioned in some early coverage. This is the vendor API's limit; third-party gateways or clients may impose lower limits.

Why consider it for visual coding and long-running tasks?

LongCat-2.5 represents images as multiple levels of discrete tokens and processes them alongside text. Its reinforcement learning tasks include real coding environments, verifiable tests and tasks reconstructed from failed agent trajectories. These designs target a cycle of understanding material, acting, reading feedback and correcting the result, rather than optimizing only a single answer. Official technical introduction

For developers, the key is feeding inspection results back to the model: does the screenshot match the target layout, do the tests pass, and can the generated files be opened? Vision support does not automatically grant browser access. The application still supplies tools, file permissions and the execution environment.

3D scenes: they must work as well as look good

The author 金色传说大聪明, writing for 赛博禅心, demonstrated a city at night, a black hole, a playable cube puzzle, a particle piano and sky lanterns, and commented on their 3D visual quality. These are third-party experiences, not independent tests by Guishu Intelligence. Source article

When reproducing similar tasks, turn acceptance criteria into checks you can actually perform:

• City scene: verify that orders and couriers really move and that pause and speed controls work; do not inspect only the lighting effects.
• Cube puzzle: check that state remains consistent after turning layers, and that scrambling and solving operate on the same state model.
• Particle piano: check consistency between touch and keyboard input, audio starting after user interaction, and usability with fewer particles.

Keep the prompt, resource budget and runtime environment fixed. Save the first result and revision history to compare the cost of completing the task across models.

Office and coding work: define what “finished” means

For business analysis, first specify fields, currency and the reporting period, then require every conclusion to be traceable to the source data. Attractive charts do not replace correct formulas. For coding, record the existing test status, restrict which directories can be changed, and ask the agent to finish with a list of modified files, test results and unresolved issues. Sending reports, releasing code or modifying production data still requires the application's own permission controls.

Reading benchmarks: compare tasks, not blanket claims of superiority

Vendor-reported evaluation	LongCat-2.0	LongCat-2.5
Terminal Bench v4.0	3.0	38.5
Program Bench Almost Resolved	2.5	25.5
AutomationBench v1.0.6	15.8	51.8

These scores come from LongCat's official evaluation table, not tests run by this site. Results depend on the tools, context budget and scoring protocol. They support an improvement in long-horizon task capability, but do not guarantee that any project will be completed in one attempt.

Pricing: distinguish vendor access from gateway multipliers

LongCat's current limited-time vendor prices per million tokens are ¥2 for uncached input, ¥0.04 for cached input and ¥8 for output. Prices can change; the vendor platform governs its billing. Official pricing

Guishu Intelligence's reference table retains a 2× multiplier, with a reference input price of ¥0.8 per million tokens. The live model catalog, checked on 2026-11-03, shows ¥0.4 input, ¥1.6 output and ¥0.04 cached input per million tokens; the input price corresponds to a 1× multiplier. The reference table and live configuration differ. Consult live pricing and actual request bills, and do not combine the two pricing systems.

When comparing long-running task costs, count all conversation turns, tool results, retries and generated output throughout the task, rather than comparing only the unit price of one input.

API integration: match the endpoint, key and model ID

The following minimal text example connects directly to the vendor, using a key issued by the LongCat platform. It is not a Guishu Intelligence endpoint. Official quick start

curl https://api.longcat.chat/openai/v1/chat/completions \
  -H "Authorization: Bearer $LONGCAT_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "LongCat-2.5",
    "messages": [{"role": "user", "content": "Explain how to validate a generated 3D scene."}],
    "max_tokens": 1024
  }'

For Guishu Intelligence, use this site's key, Base URL and corresponding protocol instead. Do not send a vendor key to another service. See the API endpoints tutorial for exact addresses and protocol details; the model page also provides current setup information.

For image requests, change content to an array of content blocks containing text and image_url. Start with one small image and a clearly defined question to inspect recognition quality, then add more images and tool calls. The vendor also documents video input, but this gateway's current catalog lists only text and images; this article does not promise gateway video support. Official visual API

Frequently asked questions

How should I choose between LongCat-2.5 and LongCat-2.0?

Evaluate 2.5 first for screenshot understanding, visual questions or modifying an interface from an image. For existing text-only workflows, compare accuracy, total cost and latency against 2.0 using the same tasks. See the LongCat-2.0 overview; do not replace every production task based only on the version number.

Does 3D support mean this is an image or video generation API?

These are different capabilities. The 3D examples here mainly involve generated code rendered by a browser or another runtime. Check the product documentation separately for a dedicated API that directly returns images or videos.

Why can a task fail even when the model is listed?

Check the model ID, the platform that issued the key, Base URL, protocol, request size and tool permissions. Long-running tasks may also hit client timeouts, output truncation or missing environment dependencies. Save the error type and request time before investigating, and avoid blindly retrying operations that may already have produced side effects.

The cover comes from 赛博禅心's LongCat-2.5 article. Model parameters and vendor evaluations follow the official sources linked above; third-party experiences and workflow recommendations are identified separately.