Skip to content
All systems operational0 AI providers monitored, polled every 2 minutes
Live status

LongCat-2.5

Flagship

by Meituan

As of November 16, 2026, LongCat-2.5 by Meituan costs $0.30 per 1M input tokens and $1.20 per 1M output tokens, with a 1M-token context window. It ranks 11th of 19 models on the TensorFeed Intelligence Index's 2026 benchmark set, with a score of 66.3 (as of November 15, 2026).

LongCat-2.5 is Meituan's successor to LongCat-2.0, and it is now the only model its pay-as-you-go API serves: the LongCat platform docs list exactly one model, LongCat-2.5, with a 1M token context window, up to 256K output tokens, and both OpenAI and Anthropic-shaped endpoints, at a limited-time discounted $0.30 per million input tokens, $0.006 on a cache hit, and $1.20 output (2 yuan, 0.04, and 8). No standard rate is published behind that discount and no end date is attached. The model card on Hugging Face describes a native multimodal mixture-of-experts with 1.6 trillion total parameters and roughly 48 billion active per token, taking image and text in and returning text, with weights under plain MIT that grant no rights to Meituan trademarks or patents, plus a W4A4 quantization at 495B for cheaper serving. Reports put the weights release at November 9 or 10, 2026, about a week after the model itself appeared; Meituan's own changelog page was unreachable here, so the exact dates are not confirmed from the vendor. The evaluation posture sits between LongCat-2.0, which shipped a technical blog and a benchmark table, and LongCat-2.5-Preview, which shipped neither: the card carries a full table but marks its unmarked figures as in-house and indicative only, and no independent evaluator has tested it. Vendor-reported highlights include Terminal Bench v2.1 at 88.4 against 89.1 for Claude Opus 5 and 88.8 for GPT-5.6 Sol, BrowseComp at 91.9, MathVision at 96.5, and Humanity's Last Exam with tools at 63.1. The infrastructure numbers are the more verifiable part, because they describe its own stack: 86 percent higher per-device throughput on long-sequence supervised fine tuning against a DDP baseline, 43.3 percent higher prefill and 42.77 percent higher decode throughput against a W8A8 MoE baseline, and 92 percent higher decode throughput for W4A4KV8 over W8A8KV16. Self-hosting targets Huawei Ascend hardware rather than NVIDIA, which continues the LongCat-2.0 story of a trillion-parameter line trained and served on domestic Chinese silicon.

Input Price

$0.30

per 1M tokens

Output Price

$1.20

per 1M tokens

Context Window

1M

tokens

Released

2026-11

Open source

Capabilities

textvisiontool-usecodereasoning

Key Strengths

  • ✓1.6T parameter multimodal MoE with roughly 48B active per token
  • ✓Plain MIT weights on Hugging Face, plus a 495B W4A4 quantization
  • ✓1M token context window with up to 256K output
  • ✓$0.30/$1.20 per 1M tokens with $0.006 cache hits, limited-time discounted
  • ✓OpenAI and Anthropic-shaped endpoints on the same API
  • ✓Terminal Bench v2.1 at 88.4, within a point of Claude Opus 5, vendor-reported

Best For

  • ▸Agentic coding and long-horizon tool use at budget pricing
  • ▸Self-hosting a trillion-parameter model under MIT
  • ▸Image plus text workloads that need a permissive license
  • ▸Deployments on Huawei Ascend rather than NVIDIA hardware

Benchmark Scores

BenchmarkScoreDescription
BrowseComp91.9Agentic web search and browsing over hard-to-find facts
Terminal-Bench 4.038.5Long-horizon agentic work in a terminal across software, science, ML, operations, hardware, security, and media (66 tasks, all-or-nothing verifiers)
Humanity's Last Exam (tools)63.1Multidisciplinary expert-level reasoning with tool access

Scores sourced from public benchmark datasets. See full benchmark leaderboard for all models.

Pricing Details

Input tokens

$0.30

per 1M tokens

Output tokens

$1.20

per 1M tokens

Estimated cost per 1K requests

$0.90

~1K input + ~500 output tokens avg

Prices are subject to change. Check the official documentation for current pricing. See the cost calculator for detailed estimates.

Open Source Model

LongCat-2.5 is free to download and self-host under the MIT. Hosted API pricing varies by provider (e.g., Together, Fireworks, Groq). See our open source LLM guide for deployment options.

Quick answers

How much does LongCat-2.5 cost?
As of November 16, 2026, LongCat-2.5 costs $0.30 per 1M input tokens and $1.20 per 1M output tokens on the Meituan API.
What is the context window of LongCat-2.5?
LongCat-2.5 has a 1M-token context window (1,000,000 tokens).
How does LongCat-2.5 rank on the TensorFeed Intelligence Index?
LongCat-2.5 ranks 11th of 19 models on the TensorFeed Intelligence Index's 2026 benchmark set, with a score of 66.3 (as of November 15, 2026). Scores from the 2024 and 2026 benchmark sets are ranked separately and never compared.

Embed the TFII badge

A live badge with this model's TensorFeed Intelligence Index rank and score. It updates with the daily index and links to the ranking.

Related Models

View DocumentationCompare ModelsCost CalculatorFull Pricing Guide