LongCat-2.5 pricing, plans and limits

Released in 2026 with 48 billion active parameters per token, LongCat-2.5 is a multimodal model aimed at teams running coding agents through Claude Code, Codex or Hermes agents. It succeeds Meituan's earlier model for API users and is a candidate for anyone testing cheaper alternatives to frontier models, with the caveat that every score so far is vendor-reported.

LongCat-2.5 is Meituan's natively multimodal mixture-of-experts model, released on its API platform on November 2, 2026. Meituan reports 88.4 on Terminal-Bench 2.1, level with the best frontier models in its own comparison table, and the model takes text, images and video inside a 1,048,576-token context window.

Provider: Meituan · Family: LongCat 2

More about Meituan on Hokai

Context window: 1,048,576 tokens · Max output: 262,144

Input modalities: text, image, video, tool-calls · Output: text, tool-calls

About LongCat-2.5

LongCat-2.5 is a natively multimodal model from Meituan for long-horizon agentic work, announced in a launch post dated November 1, 2026 and available on the LongCat API Platform on November 2, 2026. It follows LongCat-2.0 (June 30, 2026) and a LongCat-2.5-Preview that added image understanding on September 25, 2026. Meituan describes it as a mixture-of-experts model with 1.6 trillion total parameters and about 48 billion active per token. Images are turned into multi-level discrete tokens by residual vector quantization of native-resolution features, so the backbone reads pictures and text as one token stream.

The agentic coding gains come from reinforcement learning on tasks synthesized from GitHub pull requests and linked issues, each checked with fail-to-pass and pass-to-pass tests, with training mixed across several agent tools. Domain experts carried over from LongCat-2.0 were folded into one model through on-policy distillation. Training ran on domestic AI accelerators with under 80 GB of memory each, and Meituan serves the model with 4-bit expert weights and activations (W4A4), reporting 43.3% higher prefill and 42.77% higher decode throughput than its 8-bit INT8 baseline.

Every score below is vendor-reported, with thinking on, temperature 1.0 and top_p 0.95. Meituan marks rival figures with an asterisk when they come from public reports and calls the rest self-run and indicative only. On Terminal Bench v2.1 (Claude Code setup, three-run average) it scores 88.4, against 89.1 for Claude Opus 5, 88.8 for GPT-5.6 Sol, 88.3 for Kimi K3 and 88.2 for GLM-5.3; LongCat-2.0 scored 70.8. Two older models in the same table, Claude Opus 4.8 and GPT-5.5, scored 74.6 and 83.2. It posts 91.9 on BrowseComp, the highest figure in Meituan's row, and 83.1 on MCP Atlas, behind Claude Opus 5 at 85.8. On visual tests it scores 85.2 on VideoMME against 90.0 for Kimi K3, and 90.2 on CharXiv-RQ with tools. No Artificial Analysis or OpenRouter page for LongCat-2.5 was live when Hokai checked on November 3, 2026, so treat each figure as unverified.

The context window is 1,048,576 tokens shared between input and output, with a max_tokens limit of 262,144 on the model parameters page, although the quick-start page still says 128K. One request can carry up to 50 images of 10 MB each, or up to 10 videos supplied by URL, each under 100 MB and 3,600 seconds. Output is text. Thinking switches on or off per request with a budget up to 262,144 tokens, and tool calling is available through the Responses API alongside the Chat Completions and Anthropic Messages formats. Meituan's changelog lists Claude Code, Codex and Hermes as compatible agent tools.

Pricing is $0.30 per million uncached input tokens, $0.006 per million cached input tokens and $1.20 per million output tokens, which Meituan calls a limited-time discount while publishing neither a list price nor an end date. A Token Pack is a fixed quota valid for 30 days, and cache hits do not draw it down. Meituan documents only its own LongCat API Platform and chat app as access routes. The launch post marks the GitHub and Hugging Face releases as coming soon without a date, while LongCat-2.0 weights were already published on Hugging Face in July 2026.

The model suits teams running coding or browsing agents who want a 1M-token window and image or video input at a low per-token price, and visual-to-code work such as turning screenshots or design references into pages. It is a weaker pick for regulated data, because the pages Hokai read publish no retention or training terms, and for anyone who needs downloadable weights today, where MiMo-V2.6-Pro ships under the MIT license. Against LongCat-2.0 in Meituan's table, Terminal Bench v4.0 rose from 3.0 to 38.5, DeepSWE v1.1 from 9.7 to 64.6, and Meituan's internal AppDev test from 47.7 to 73.7. Hokai's model access tracker shows how each model can be used, and the Meituan models page lists Meituan's models.

Pricing

These are the pay-as-you-go prices on the international site on November 3, 2026, described by Meituan as a limited-time launch discount; no list price or end date is published, and the China site prices in yuan. The pricing page lists no free tier. Enterprise-verified accounts get tiered discounts on monthly spend. Account rules for the platform are covered on the [Meituan](/hub/companies/meituan) record.

What a real job costs

Budgets: 20-page PDF = 30k in / 1k out · Support reply = 2k in / 300 out · Coding agent run = 200k in / 20k out. Computed from the vendor's per-token prices at render time; cached-input discounts are not applied.

Key Features

• Image and video input: Accepts as many as 50 images, or 10 videos given as URLs, per call; image files may reach 10 MB and video files 100 MB.
• Thinking on or off per request: A request-level switch with a thinking budget of up to 262,144 tokens, returned as reasoning_content.
• Three API formats: OpenAI Chat Completions, OpenAI Responses with multi-turn tool calling, and Anthropic Messages all accept the same API key.
• 4-bit serving: Meituan runs expert weights and activations at 4 bits and reports 43.3% higher prefill throughput than its 8-bit baseline, a vendor-reported figure.

Pros

• Meituan publishes tool versions, sampling settings and which rival scores are third-party, so its table is easier to audit than most launch posts.
• Existing OpenAI and Anthropic SDKs connect by changing the base URL and model name, so adopting it needs little rework.
• Visual input lives in the model that writes the code, which suits fixing front-end bugs from screenshots.

Cons

• Every score is vendor-reported, and no independent leaderboard listed it on November 3, 2026.
• It trails the best listed models on deeper repository and terminal tasks in Meituan's own table.
• The pages checked state no retention, training-on-data or SLA terms, and the weights are only promised.
• Launch pricing is flagged as limited-time with no published list price, so budgeting past the discount is guesswork.

Benchmarks

• MCP Atlas: 83.1% vendor-reported · 03 Nov 2026 — Tasks solved by calling external tools over the Model Context Protocol, % solved.
• Terminal-Bench 2.1: 88.4% vendor-reported · 03 Nov 2026 — Multi-step tasks completed in a real command line, % solved.

A benchmark is an exam, not the job. Scores transfer unevenly between tasks, so weigh the one closest to your workload and read every figure with its source.

Visit LongCat-2.5 Official Page