Before You Call LongCat-2.5, Check the Model Name and Max Output Docs

Meituan’s LongCat API open platform Changelog dated 2026-11-02 lists “LongCat-2.5 上线”—the GA release of LongCat-2.5. That entry sits beside a separate 2026-09-25 line for LongCat-2.5-Preview. Teams that already pointed harnesses or SDKs at Preview need a short pre-flight: the model id string, which base URL the client uses, what the docs say about max output, and how the current usage-based billing table is labeled. Mixing Preview and GA into one “LongCat-2.5 just shipped” story is how wrong model values and wrong output caps land in production configs.

Before cutover, line up the model id, base URL, max-output figures, and the usage table against the LongCat API Changelog, Quick Start, model-parameters reference, and LongCat-2.5 pricing pages. We haven't run every OpenAI-SDK, Anthropic-SDK, Claude Code, Codex, or Hermes path against the live endpoints yet; prior live reports can go through the site-owner contact at the bottom.

Related on-site context: Volcengine Ark Coding Plan Softens Deduction Rates for Kimi and DeepSeek Models, DeepSeek Tool Calls Return 400 When reasoning_content Is Dropped, OpenAI API Pricing (October 2026), and Mistral Puts Responses API in Public Preview for Agents.

What the Changelog lists on 2026-11-02 versus 2026-09-25

Under 版本: 2026-11-02, the Changelog titles the GA item LongCat-2.5 上线 and summarizes three themes:

1. Native multimodal work that pairs visual understanding with agent-style coding (visual programming, creative design, and 3D-oriented tasks in the Changelog wording).
2. Long-horizon agents for complex codebases and full projects, including multi-step reasoning, tool use, and end-to-end execution.
3. Harness fit: compatibility with Claude Code, Codex, Hermes, and similar mainstream harnesses so existing workflows can plug in as needed.

Under 版本: 2026-09-25, the Changelog keeps a distinct LongCat-2.5-Preview 上线 entry (multimodal understanding, coding focus, and deep adaptation to Claude Code–class environments). Preview and GA stay two dated rows—not one soft rename. A client that still sends a Preview-era model string still needs a check against the live supported-models table on Quick Start before assuming GA traffic is already flowing.

The 2026-11-02 GA entry does not publish a Hugging Face or GitHub open-source address for LongCat-2.5. Older Flash-family Changelog rows do list open-source links; those rows do not mean 2.5 GA is open-sourced.

Model name and dual endpoints

Quick Start’s supported-models table lists a single GA row for this wave:

The string to put in model is LongCat-2.5—same spelling for both API shapes in the Quick Start samples.

Base URLs (Quick Start “接入端点”):

• OpenAI-compatible: https://api.longcat.chat/openai
• Anthropic-compatible: https://api.longcat.chat/anthropic

Interface surfaces called out on Quick Start:

• OpenAI shape: chat completions at /v1/chat/completions
• Anthropic shape: messages at /v1/messages

SDK sketches on Quick Start set base_url (OpenAI Python SDK) or base_url plus default headers (Anthropic Python SDK) to those hosts, with model="LongCat-2.5". cURL samples hit …/openai/v1/chat/completions and …/anthropic/v1/messages with a Bearer API key. Auth and key creation live on the platform’s API Keys flow; account signup follows the platform’s own signup steps.

An OpenAI-compatible client aimed at the Anthropic host—or the reverse—is a different failure mode from a wrong model string. Host and path family are the first checks before timeout or 404 noise.

Context window versus max output—and the 128K / 256K doc split

Two numbers agree across Quick Start and model-params: the context window for LongCat-2.5 is 1M tokens. Model-params spells that as 1,048,576 (1M) for the total tokens a single request can hold (input plus output, per that page’s definition).

Max output is where the documentation disagrees today:

Model-params further states that max_tokens caps generated tokens only (input not included). Hitting the cap truncates generation and sets finish_reason to length. Values outside the documented ranges for top_p, top_k, repetition_penalty, and max_tokens return 400 with invalid_request_error / invalid_parameter.

Neither 128K nor 262,144 / 256K is crowned the single correct ceiling here. Both figures appear on the official documentation pages. Integration code and product copy keep both figures visible until Quick Start and model-params converge—or until a live call’s accepted max_tokens and observed truncation behavior are recorded for that account. We haven't measured which ceiling the runtime enforces on either endpoint. For reference only.

Practical reading for agent and coding harness configs:

1. 1M context is the shared headline capacity.
2. Max output stays documentation-inconsistent: Quick Start 128K versus model-params max_tokens 262,144 (256K).
3. Raising max_tokens to 256K in a harness “because the params table says so” without also noting the Quick Start 128K line—or the reverse—hides the split.

Temperature, thinking switch, and thinking_budget

Model-params fixes temperature to the thinking mode rather than leaving it freely tunable:

• Thinking on → temperature fixed at 1.0
• Thinking off → temperature fixed at 0.7

Passing another temperature value does not raise an error in the documented behavior; the platform overwrites it with the fixed value for the current thinking state. Clients that assume “temperature 0.2 stuck” will misread logs if thinking mode flipped underneath.

Thinking is controlled with a thinking object:

• {"type":"enabled"} — thinking on
• {"type":"disabled"} — thinking off

thinking_budget is an integer ceiling for thinking tokens, maximum 262,144. Other sampling knobs on the same page include top_p (default 0.95, range (0, 1]), top_k (default -1, range [-1, 100]), and repetition_penalty (default 1, range [1, 1.5]). Those details matter when porting a DeepSeek- or OpenAI-style agent loop; still read the live params table after a docs update.

For a parallel class of “multi-turn agent request rejected because a reasoning field was dropped,” see the on-site DeepSeek tool-call write-up linked above—that failure mode is DeepSeek-specific, but the habit of preserving vendor reasoning fields across tool rounds transfers cleanly when evaluating any thinking-capable API.

Limited-time pricing—no end date on the page

The LongCat-2.5 pricing page states that current API usage-based billing supports LongCat-2.5. Table headers read 折扣价 / 百万 Tokens（限时）—discounted price per million tokens, limited-time. A page note adds that the discount is a launch-period limited offer and that actual charges follow settlement.

Figures on the pricing page (per million tokens):

No end date appeared on that pricing page in the same check. These rows are a poor fit for a permanent list-price spreadsheet, an invented sunset day, or a same-day reading of “limited-time.” When budgeting agents or coding harnesses, the pricing page and the account settlement view are the live sources; if the table or the limited-time label changes, the live page wins. For broader API price-wall habits (token vs subscription, cache vs output lines), the on-site OpenAI API Pricing (October 2026) guide is a separate vendor wall; use LongCat’s own pricing page for LongCat rates.

We haven't reconciled a live LongCat invoice line-by-line against these published cells. For reference only.

Harness and agent wiring notes

Changelog GA copy calls out compatibility with Claude Code, Codex, Hermes, and similar harnesses. The Changelog states compatibility; we haven’t measured pass rates. When wiring a harness:

• Provider base URL matches the LongCat host (OpenAI-shaped vs Anthropic-shaped).
• model is LongCat-2.5 unless the platform later lists aliases.
• max_tokens / output caps follow the dual documentation figures above; which page the team followed belongs in the runbook.
• Temperature is overridden by thinking mode; “temperature ignored” is an SDK bug only after thinking state is logged.
• Thinking-budget and tool-round message history stay consistent with the vendor’s rules for the chosen API shape.

Volcengine Ark coding-plan deduction changes (linked above) and Mistral’s Responses API preview (also linked) are different vendors’ surfaces—useful when a shop runs several agent backends side by side, not substitutes for LongCat’s own Quick Start samples.

Pre-flight checklist

1. Release dates stay split. GA Changelog row = 2026-11-02; Preview row = 2026-09-25—two rows, not one launch.
2. Model string. LongCat-2.5 on both OpenAI- and Anthropic-compatible calls per Quick Start.
3. One endpoint family, matching paths. https://api.longcat.chat/openai vs https://api.longcat.chat/anthropic.
4. Both max-output figures stay cited. Quick Start 128K; model-params max_tokens 262,144 (256K). Internal runbooks keep both numbers.
5. Context = 1M / 1,048,576 as documented; max_tokens counts output only.
6. Temperature is fixed by thinking mode (1.0 on / 0.7 off); other values are overwritten without error.
7. thinking = enabled or disabled; thinking_budget max 262,144.
8. Pricing is labeled limited-time with no published end date on the page in this check—the live table is the source before any cost model hardens.
9. Older Flash Changelog rows do not mean 2.5 GA is open-sourced.
10. Open the four documentation URLs again before cutover; the docs can still change after an internal runbook is frozen.

Bottom line

LongCat-2.5 is listed as GA on Meituan’s LongCat API open platform under the 2026-11-02 Changelog entry, distinct from LongCat-2.5-Preview on 2026-09-25. Calls use model id LongCat-2.5 on either https://api.longcat.chat/openai or https://api.longcat.chat/anthropic, with the 128K vs 262,144 (256K) max-output documentation split kept visible in configs, and read the pricing table’s limited-time label from the live page; don’t treat it as a permanent rate card or invent a deadline. Every number on the Changelog, Quick Start, model-params, and pricing pages still needs a fresh check before Preview traffic moves.