Meituan launched LongCat-2.5 on November 2, a 1.6-trillion-parameter multimodal model for long-running AI agents, trained entirely on Chinese chips. Its API lists at $0.30 per million input tokens and $1.20 per million output tokens, the exact peak price of DeepSeek V4.1 Flash and up to 20x cheaper than Claude Opus 5. Meituan’s own benchmarks put it near Opus 5, not above it.
Key takeaways
- LongCat-2.5 costs $1.20 per million output tokens versus $25 for Claude Opus 5.
- The model has 1.6 trillion total parameters, with about 48 billion active per token.
- On Terminal Bench v4.0 it scores 38.5, well behind Claude Opus 5’s 53.9.
What is LongCat-2.5?
LongCat-2.5 is Meituan’s new flagship AI model, built for agents that run long coding and browsing jobs. It reads text and images in one framework, supports a 1M-token context window, and is live in Meituan’s chat app and API. It replaces LongCat-2.0, which Meituan released in June.
Meituan is best known as China’s food-delivery and local-services giant. It is now shipping frontier-scale models. TechNode reported that LongCat-2.5 has 1.6 trillion total parameters, with about 48 billion activated per token.
That is a mixture-of-experts design. Only a small slice of the network runs on each token, which is what keeps serving costs low enough to price the model like a budget tier.
Trained on domestic chips
Meituan says training ran end to end on Chinese-made accelerators. The LongCat launch post states: “We optimize long-context training and W4A4 inference on domestic AI accelerators.” It does not name the chip vendor.
According to Meituan’s announcement, its training setup lifted per-card throughput 86% over a standard DDP baseline. Its 4-bit W4A4 inference ran about 40% faster than an 8-bit baseline in stress tests.
How much does LongCat-2.5 cost?
LongCat-2.5 costs $0.30 per million uncached input tokens, $0.006 per million cached input tokens, and $1.20 per million output tokens. Meituan labels these “limited-time” discounted prices and gives no end date. They match DeepSeek’s peak rates for V4.1 Flash to the tenth of a cent.
The LongCat pricing page lists the same rates in yuan: ¥2 input, ¥0.04 cached, and ¥8 output. It adds that “Prices are subject to change.”
The match with DeepSeek is no accident. The DeepSeek pricing page lists V4.1 Flash at exactly $0.30, $0.006 and $1.20 during peak hours. DeepSeek halves those rates off-peak, which LongCat does not.
LongCat-2.5 pricing vs. rivals
The table below uses each provider’s list price and a sample agent workload of 10 million uncached input tokens and 2 million output tokens.
| Model | Input / 1M | Output / 1M | 10M in + 2M out |
|---|---|---|---|
| LongCat-2.5 (launch discount) | $0.30 | $1.20 | $5.40 |
| DeepSeek V4.1 Flash (peak) | $0.30 | $1.20 | $5.40 |
| DeepSeek V4.1 Flash (off-peak) | $0.15 | $0.60 | $2.70 |
| DeepSeek V4 Pro (peak) | $1.32 | $3.96 | $21.12 |
| Claude Sonnet 5.5 | $2.00 | $10.00 | $40.00 |
| Claude Opus 5.5 | $4.00 | $20.00 | $80.00 |
| GPT-5.6 Sol (prompts under 272K) | $4.00 | $20.00 | $80.00 |
| Claude Opus 5 | $5.00 | $25.00 | $100.00 |
On that workload, Claude Opus 5 costs about 18.5 times more than LongCat-2.5. Anthropic’s pricing page lists Opus 5 at $5 input and $25 output.
OpenAI’s GPT-5.6 Sol model page lists $4 input and $20 output. Prompts over 272K tokens are billed at 2x input and 1.5x output, so long-context agent runs cost more than the table shows.
Is LongCat-2.5 as good as Claude Opus 5?
Not across the board. Meituan says LongCat-2.5 scores between Claude Opus 4.8 and Claude Opus 5 on most mainstream benchmarks and beats Opus 5 on some. Its own table shows a lead in web browsing and a clear gap in hard terminal and software-engineering tasks. Every score so far comes from Meituan.
| Benchmark | LongCat-2.5 | Claude Opus 5 | GPT-5.6 Sol |
|---|---|---|---|
| Terminal Bench v4.0 | 38.5 | 53.9 | 37.3 |
| DeepSWE v1.1 | 64.6 | 68.8 | 72.7 |
| BrowseComp | 91.9 | 90.8 | 90.4 |
| MCP Atlas | 83.1 | 85.8 | 81.8 |
| JobBench | 65.4 | 66.7 | 45.4 |
The jump from the last version is real on paper. Terminal Bench v4.0 rose from 3.0 on LongCat-2.0 to 38.5, and DeepSWE v1.1 rose from 9.7 to 64.6, according to the same launch data.
The comparison set is a generation old
Here is the skeptical read. Meituan benchmarks against Claude Opus 5 and GPT-5.6 Sol, not the current flagships. Anthropic shipped Opus 5.5 in September, and OpenAI and Google have since released GPT-6-class and Gemini 4 models.
Rival scores in Meituan’s table also come from mixed sources: public leaderboards and rivals’ own reports. Until independent labs test LongCat-2.5 on the same harness, “near Opus 5” is a vendor claim, not a fact.
How can developers use LongCat-2.5?
Developers can call LongCat-2.5 today through Meituan’s API, which offers OpenAI-format and Anthropic-format endpoints. Consumers can try it at longcat.ai. Open weights on GitHub and Hugging Face are marked “coming soon,” with no date. That matters, because LongCat-2.0’s weights shipped under the MIT license.
Practical details worth knowing before you switch:
- Context: 1M tokens in, up to 128K tokens out, per the LongCat API docs.
- Claude Code: Meituan publishes a setup guide, but a developer write-up notes Claude Code assumes a 200K window for the unknown model ID unless you override it.
- Caching: cached input costs $0.006 per million tokens, 50 times less than uncached input, so repeated agent context gets very cheap.
- Data location: the API runs on Meituan’s platform in China, which many US and EU enterprises will rule out on compliance grounds.
Who wins and who loses from LongCat-2.5?
Cost-sensitive agent builders win: another near-frontier model at $1.20 output keeps pressure on prices. DeepSeek loses its clean price lead. Anthropic and OpenAI keep their enterprise moat for now, but every Chinese release at this price narrows the premium they can charge for routine agent work.
Winners
Developers and startups. A 10M-in, 2M-out workload at $5.40 instead of $80 changes unit economics for any product that runs long agent loops. We saw the same squeeze when Kimi K3.1 Pro launched at half the price of Opus.
China’s chip makers. A 1.6-trillion-parameter model trained only on domestic accelerators is a sales pitch for local silicon, even without a named vendor.
Meituan’s core business. CEO Wang Xing said on the August 28 earnings call that the company “will not become a token factory,” per Meituan’s announcement. The models feed tools like Miaoshou, an AI workbench for local merchants.
Losers
DeepSeek. Its V4.1 Flash 1017 update was built around cheap per-task cost. Meituan now matches that peak price with a far larger model. DeepSeek’s off-peak discount is its remaining edge.
Western mid-tier pricing. OpenAI already cut prices hard with GPT-6.1 Terra at $1 per million. Releases like LongCat-2.5 make it harder for any lab to hold $10-plus output pricing on tasks where a cheaper model is good enough.
There is a tension in Meituan’s strategy. A company that says it will not be a “token factory” is selling tokens at a discount it calls temporary. Expect the discount to end once the model has users.
What to watch next
Three signals will decide whether LongCat-2.5 is a real price-setter or a launch promotion: the open-weights release, the end of the discount, and independent benchmarks. One dated event also matters for the price comparison above.
- November 21, 2026: OpenAI says GPT-5.6 Sol’s promotional pricing runs at least through this date. Any change shifts the gap with LongCat.
- Hugging Face and GitHub release: marked “coming soon.” Watch the license; LongCat-2.0 used MIT.
- End of the “limited-time” price: Meituan gives no date. A move above $1.20 output would end the DeepSeek parity.
- Third-party scores: results from independent leaderboards will show whether the Opus 5 comparison holds.
FAQ
When was LongCat-2.5 released?
Meituan announced LongCat-2.5 on November 2, 2026. The LongCat team’s blog post is dated November 1.
How much does the LongCat-2.5 API cost?
$0.30 per million uncached input tokens, $0.006 per million cached input tokens and $1.20 per million output tokens, at a limited-time discount.
Is LongCat-2.5 open source?
Not yet. GitHub and Hugging Face releases are listed as coming soon with no date. The earlier LongCat-2.0 weights are public under the MIT license.
How big is LongCat-2.5?
It has 1.6 trillion total parameters, with about 48 billion active per token. It supports a 1M-token context window and up to 128K output tokens.
Does LongCat-2.5 beat Claude Opus 5?
On some tests. Meituan reports 91.9 vs 90.8 on BrowseComp, but 38.5 vs 53.9 on Terminal Bench v4.0. No independent results are out yet.
What chips was LongCat-2.5 trained on?
Meituan says training ran entirely on domestic Chinese AI accelerators but has not named the vendor.
The bottom line
LongCat-2.5 is a pricing event more than a capability event. It does not beat today’s frontier, and its benchmarks are self-reported against last-generation rivals. But it puts a 1.6-trillion-parameter agent model at DeepSeek’s exact price, trained entirely on domestic Chinese chips.
Verdict: worth testing for browsing-heavy and high-volume agent jobs where data location is not a blocker. Not a replacement for Opus or GPT on hard coding. And budget for the “limited-time” price to rise.
Free guide · 12 pages
AI Tools Every Investor Should Use in 2026
10 tools for research, stock signals, charts and crypto, what each one costs, three ready-made stacks from $0 a month and 7 copy-paste prompts for filings and earnings calls. Enter your email and we will send it to you.
Plus one email a week: AI Money This Week, every Sunday. Unsubscribe anytime.
Sources
- Meituan: LongCat-2.5 announcement (Economic Observer), Nov 2, 2026
- LongCat team: Introducing LongCat-2.5
- LongCat API Platform: LongCat-2.5 pricing
- TechNode: Meituan launches LongCat-2.5, Nov 2, 2026
- APIYI: LongCat 2.5 benchmark compilation
- DeepSeek API: models and pricing
- Anthropic: Claude API pricing
- OpenAI: GPT-5.6 Sol model page
- DEV Community: LongCat-2.5 in Claude Code
Wealth Engine researches and drafts with AI tools and checks every figure against the sources above. How we report.
Leave a Reply