LongCat 2.5 Preview API: Image Inputs, Coding and Integration Guide

LongCat 2.5 Preview API access gives developers a route to Meituan's newer coding and multimodal model. The integration decision starts with the endpoint and client you plan to use. A model's image-understanding capability, its native API contract and a coding client's exposed features are separate things to verify.

LongCat's platform change log records the preview's availability, while the native chat API reference supplies the exact identifier and request format. The examples below describe a controlled pilot rather than a completed migration.

LongCat 2.5 Preview API routes and model IDs

The native model identifier is LongCat-2.5-Preview. Its OpenAI-compatible base URL is https://api.longcat.chat/openai/v1. The platform also documents an Anthropic-format route; copying a base URL between protocols can create an incorrect request path.

A gateway can use a different identifier. For example, ZenMux's model listing names the route meituan/longcat-2.5-preview. Keep the selected provider's model name with its endpoint instead of mixing native and gateway values.

The output setting is not the context window. LongCat's public site describes a million-token long context, but the input budget served through your chosen route and the client's limits still need checking. Do not fill the window simply because a large number appears on a product page.

Send a short native API request

Use a platform key with the intended account quota. A short text request can establish authentication and model selection before adding tools or images:

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-Preview",
    "messages":[{"role":"user","content":"Explain why a cache key should include the model version."}],
    "max_tokens":1024,
    "thinking":{"type":"disabled"}
  }'

This example uses the documented thinking field and a bounded output. It does not add image content to a text-only request shape or claim that a URL pasted into text becomes a decoded image.

Save the response model, usage, finish reason and HTTP status. If authentication or quota fails, resolve the account condition before changing prompts. For a gateway, repeat the baseline using that gateway's documented credentials and model identifier.

Streaming is documented as an optional SSE response. Test non-streaming first if your client makes it difficult to inspect complete responses, then verify that streaming preserves the final usage and error handling your application needs.

Separate image capability from client support

The platform release notes describe image understanding for cross-modal questions, summaries and visual reasoning. That is a useful capability statement, but a working text call does not establish the image input schema for every route.

The reviewed general chat reference describes content as a plain-text string. For an image pilot, consult the selected route's current multimodal contract or a documented integration that exposes image attachments. Do not invent an image field from another vendor's request example.

Use an image with a verifiable answer, such as a small diagram with labeled relationships, and inspect what the client actually sends. Then compare a second request without the image to determine whether the answer depends on the visual evidence.

Test readability and failure cases as well as a clean sample. A client might resize an image, omit it from a later turn or convert it into extracted text. Those behaviors change what the model receives and should be recorded separately from the model's advertised capability.

Configure thinking and coding tools

The official OpenCode integration guide describes a built-in LongCat provider. Its documented flow is to run opencode auth login, select LongCat and supply the platform key. If the provider is absent, check the client version before building a custom configuration around an older interface.

After authentication, select the preview model through the client's model picker. The guide also documents configuring the thinking field in model options, with thinking enabled by default. Confirm that the option reaches the request rather than only changing the interface label.

Start a repository pilot with a read-only task, then allow a narrow edit on a disposable copy. Tool support requires the client, model response parser and executor to agree. Inspect both the proposed arguments and the actual tool result.

Use a known failing test as the task's destination. Judge the resulting patch and command output, including unrelated changes and repeated failures. Do not treat a confident completion message as proof that the test ran.

Evaluate long-context work and billing

A long-context coding task should need the supplied context. Begin with a focused issue and relevant files, then increase the surrounding material when it changes the answer. Record missing details, incorrect cross-file assumptions and whether information near the end of the input is used correctly.

Measure time to a useful answer and total task completion separately. Thinking, context processing and tool execution can all affect the session. For general background, our long-context comparison guide explains why capacity alone is an incomplete selection rule.

Keep native API quota, token packs and gateway billing separate. Account plans can have different balances or conditions, and a promotion on one route does not establish another route's price. Use the account's current pricing and usage record for the actual pilot.

Compare accepted tasks across routes, including retries and fallbacks. A successful native request and a successful gateway request can still differ in latency, supported fields and failure recovery. Record those differences before choosing the application's default provider.

Conclusion: migrate only the verified LongCat workflow

LongCat 2.5 Preview API evaluation should begin with the exact route and model identifier. Establish text access, verify thinking controls and test the tools or image path the application actually uses. Expand context gradually and compare complete accepted tasks rather than assuming that advertised capabilities reach every client unchanged.

Frequently asked questions

Is the native model ID the same as a gateway ID?

Not necessarily. The native identifier is LongCat-2.5-Preview; use the selected gateway's documented identifier for its endpoint.

Can I add an image URL to any text request?

A URL in text is not proof that the model received an image. Follow the route's documented image-input contract and inspect the client request.

Does turning off visible reasoning disable thinking?

A display setting may not change the request. Verify the documented thinking field is forwarded by the integration.