Meituan LongCat-2.5: 1.6T Params at One-Fiftieth the Price

A 1.6-trillion-parameter Chinese model just matched the US frontier on terminal work

On November 1, 2026, Meituan's LongCat team shipped LongCat-2.5 — a native multimodal mixture-of-experts model with 1.6 trillion total parameters and roughly 48 billion activated per token. Text, image and video go in; text comes out. It follows a preview that appeared on the LongCat API platform on September 25, which shipped without a published benchmark table. This release comes with the numbers.

The headline is not the parameter count. It is that Meituan trained and serves this model entirely on domestic Chinese accelerators, and that on its own evaluation table it beats several US frontier models on specific agentic coding tasks — while pricing input at $0.30 and output at $1.20 per million tokens.

For comparison, Claude-Opus-5 appears in the same table at $15 input / $75 output per million. That is a 50x gap on input and 62.5x on output.

What the official evaluation table actually shows

Every figure below is from Meituan's launch post. Scores marked with an asterisk in the original are cited from third-party reports or public leaderboards; the rest are either self-run by Meituan or from the developer's own materials. Read them with that distinction in mind.

Where LongCat-2.5 lands ahead or level:

• Terminal Bench v2.1: 88.4. GPT-5.5 scores 83.2 and Claude-Opus-4.8 scores 74.6. LongCat-2.5 sits in the same band as GPT-5.6-Sol (88.8), Claude-Opus-5 (89.1) and Kimi K3 (88.3)
• Terminal Bench v4.0: 38.5. The previous generation scored 3.0. Claude-Opus-4.8 scores 23.6, Claude-Opus-5 leads at 53.9
• DeepSWE v1.1: 64.6, up from 9.7 — a 6.7x jump. GPT-5.5 is at 67.0
• AutomationBench v1.0.6: 51.8, ahead of GPT-5.5 at 39.7 and Claude-Opus-5 at 50.3
• BrowseComp: 91.9, ahead of GPT-5.5 at 84.4 and Claude-Opus-5 at 90.8, level with Kimi K3 (91.2) and DeepSeek-V4.1-Flash (91.5)
• JobBench: 65.4 against GPT-5.5 at 38.3 and Claude-Opus-4.8 at 48.4

And where it clearly loses:

• NL2Repo: 60.5 against Claude-Opus-5 at 75.3
• SWE Atlas Test Writing: 47.8 against 62.2
• FORTE workflow simulation: 83.3, behind GLM-5.3 (85.6) and Kimi K3 (86.3)
• VideoMME: 85.2, behind GPT-5.6-Sol (89.5) and Kimi K3 (90.0)

The honest summary: LongCat-2.5 has closed the gap on terminal-driven and browser-driven agent work. It has not closed the gap on repository-scale generation or test authoring. Claude-Opus-5 still wins overall.

The engineering that makes 1.6T parameters affordable

Three implementation choices from the technical blog matter more than the marketing framing.

Images are discretized into tokens, not passed through a separate vision encoder. Visual signals at native resolution are encoded into multi-level vocabulary tokens via residual vector quantization over features from a semantic-and-aligned encoder, following the LongCat-Next approach. Because a fixed tokenizer lets image token IDs be precomputed and cached offline, the backbone only ever sees pretokenized sequences. That removes the vision encoder from the training graph, which simplifies long-sequence packing and keeps modality load roughly uniform across data-parallel ranks. Meituan validated the shared semantic space only on a smaller ~3B-active MoE, so treat the "unified embedding space" claim as preliminary. W4A4 inference on native INT4 units. Both expert weights and activations are quantized to 4 bits and the matrix multiplications run on the accelerators' INT4 pipelines, building on FlatQuant's learnable equivalent transformations. Against a W8A8 INT8 baseline, Meituan reports prefill throughput up 43.3%, decode throughput up 42.77%, time-to-first-token down 15.91%, time-per-output-token down 19.72%, and an average 1.73x speedup on the grouped matrix multiply kernel. The company claims no measurable downstream accuracy loss, which is the claim most worth verifying yourself. Dual-shard FSDP for sub-80GB accelerators. Parameter FSDP sharding stays inside each supernode, while FP32 master weights and optimizer states are further sharded across FSDP domains through HSDP replicas. Hybrid vocabulary parallelism shards the embedding layer, LM head and N-gram embeddings across the DP × CP domain. The reported result is 86% higher per-device throughput than a DDP baseline on long-sequence SFT.

The pattern here is a hardware-constrained answer: Chinese accelerators have less memory bandwidth per device, so the model compensates with aggressive quantization and topology-aware parallelism rather than raw hardware headroom. That is exactly the part that determines whether the listed price is sustainable.

What this means if you are evaluating it

Model the price gap against your own volume. At $0.30/M input, $1.20/M output and $0.006/M on cache hits, a workload of 10 million tokens a month (split evenly) costs roughly $7.50 on LongCat-2.5 and roughly $450 on Claude-Opus-5. If your pipeline is token-heavy and quality-tolerant — bulk summarization, log triage, first-pass triage agents — the arithmetic is not close. Do not treat the vendor table as a proxy for your tasks. Meituan chose the harness: Claude Code v2.1.280 with a 512K max sequence length for Terminal Bench, mini-swe-agent at 1M for DeepSWE, Codex harness for SkillsBench. Terminal Bench v2.1 is averaged over three independent runs, which is more rigorous than most launch posts. But harness choice moves rankings by several points. Run 20 to 50 of your own repository tasks against both models before committing — the LongCat side is cheap enough that this step costs almost nothing. Start with a model-ID swap, not a rewrite. The LongCat API documentation lists explicit compatibility with Claude Code, Hermes, OpenClaw, OpenCode and Kilo Code, and the endpoint accepts both OpenAI-compatible and Anthropic-compatible request formats. That makes a side-by-side trial a configuration change rather than a migration project.

The weights are not open

This is easy to miss. The launch post links GitHub and Hugging Face as "Coming soon" — no weights were published with this release. LongCat-2.0, by contrast, was open-sourced alongside inference code for domestic accelerators. LongCat-2.5 is API-first.

So if self-hosting is a hard requirement — data residency, air-gapped deployment, fine-tuning — this model is not currently an option for you regardless of price. It belongs in the hosted-endpoint comparison alongside GPT-6.1 Sol and Gemini 4 Carbon, not in the open-weight conversation with DeepSeek or Kimi K3.

Bottom line

The accurate reading is narrower than the release headline. LongCat-2.5 has reached parity with the US frontier on terminal operation, browser-based research and office-task automation, at roughly one-fiftieth the per-token cost, trained on non-NVIDIA silicon. It has not reached parity on repository-scale code generation or test writing.

That combination argues for splitting workloads rather than switching providers. Put cost-dominated batch work on LongCat-2.5 and keep quality-critical development on Claude-Opus-5 or GPT-6.1 Sol. Measure the failure rate on your own tasks for a week before you move anything that has a customer attached to it.