Meituan's LongCat-2.5 Is a 1.6-Trillion-Parameter MoE With About 48 Billion Active. The $0.30 and $1.20 API Price Is Limited-Time, With No End Date.
Meituan's LongCat blog of November 1 describes LongCat-2.5 as a native multimodal model with 1.6 trillion total parameters and approximately 48 billion activated per token. The API docs list a 1,048,576-token context and a 262,144-token output limit, and the pricing page lists a limited-time rate of $0.30 per million input tokens, $0.006 cached, and $1.20 output, with no end date and a note that prices are subject to change. Temperature is fixed by thinking state. The benchmark table is the vendor's: Terminal Bench v4.0 at 38.5 against 37.3 for GPT-5.6 Sol and 53.9 for Claude Opus 5, with the rival figures footnoted as taken from public reports. No open weights for 2.5 were announced. We read the blog and the API docs.
By T.W. Ghost
Meituan's LongCat team published LongCat-2.5 today, November 1, 2026: a native multimodal model that the blog describes as "1.6 trillion total parameters and approximately 48 billion activated per token." The API documentation puts the context window at 1,048,576 tokens and the maximum output at 262,144. The pricing page lists $0.30 per million input tokens, $0.006 for cached input, and $1.20 per million output, or 2, 0.04, and 8 yuan, under a limited-time label with no end date and a line that prices are subject to change.
Everything about the model's size is the vendor's number, and so is every benchmark. The blog's table puts LongCat-2.5 at 38.5 on Terminal Bench v4.0, above GPT-5.6 Sol at 37.3 and well below Claude Opus 5 at 53.9, with the rival figures marked as taken from public reports and leaderboards rather than run by Meituan. One API detail will catch people: temperature is fixed, 1.0 with thinking on and 0.7 with thinking off, and other values are overridden. There is no open-weight release for 2.5 on the blog, and no independent score. We read the blog, the pricing page, and the model-parameter docs. Here is what held up.
What You Are Calling
| LongCat-2.5 | |
|---|---|
| Published | November 1, 2026 |
| Size | 1.6T total parameters; about 48B active per token (vendor) |
| Modality | Native multimodal (vendor) |
| Context | 1,048,576 tokens (API docs) |
| Max output | 262,144 tokens (API docs) |
| Price, limited-time | $0.30 in / $0.006 cached / $1.20 out per 1M tokens; also 2 / 0.04 / 8 yuan; no end date; "prices are subject to change" |
| Temperature | Fixed: 1.0 with thinking on, 0.7 with thinking off; other values overridden |
| Open weights | Not announced for 2.5 |
| Independent score | None found |
A limited-time price with no limit printed is the same pattern as Mercury 3's launch rate: budget at a number the vendor has not published, or screenshot the page with a date.
Forced temperature breaks assumptions. Samplers, eval harnesses, and routing layers that set temperature per request will silently get 1.0 or 0.7 instead. The docs say so.
The second trillion-parameter MoE of the week. Mistral Large 4 shipped weights at 1.05T total and 52B active three days earlier, with a license. LongCat-2.5 is larger on paper and closed in practice.
The Benchmarks, All of Them Meituan's Table
| Terminal Bench v4.0 (LongCat blog) | Score |
|---|---|
| LongCat-2.5 | 38.5 |
| GPT-5.6 Sol | 37.3, from public reports per the blog footnote |
| Claude Opus 5 | 53.9, from public reports per the blog footnote |
The win is over a mid-tier model from the previous generation by just over a point, on the vendor's own run; the loss is to a frontier model by more than fifteen, on numbers the vendor copied from elsewhere. That is an honest table, and it does not say "frontier."
Verified vs Unconfirmed: The Scorecard
| Claim | Verdict |
|---|---|
| 1.6T total, about 48B active, native multimodal | Vendor (LongCat blog) |
| Context 1,048,576; max output 262,144 | Verified (LongCat API docs) |
| $0.30 / $0.006 / $1.20 limited-time; prices subject to change | Verified as vendor promo (pricing page) |
| Temperature fixed by thinking state | Verified (model-params docs) |
| Terminal Bench v4.0 38.5 vs 37.3 and 53.9 | Vendor (blog table; rivals from public reports) |
| Open weights for 2.5 | Unverified (not announced) |
| Independent benchmark | Unverified (none) |
Who Should Use It
Try LongCat-2.5 if a million-token context at thirty cents per million input tokens fits a workload you can evaluate yourself, and your stack does not depend on setting temperature.
Skip the size and the table as reasons to switch; both are Meituan's, the price has no floor date, and nobody independent has scored it.
The frontier is elsewhere: on the vendor's own chart Opus 5 is fifteen points up, and on the rebuilt Artificial Analysis index LongCat-2.5 does not appear at all. Our model comparison has the current standings side by side.
Sources
- LongCat blog: LongCat-2.5
- LongCat platform docs: LongCat-2.5 pricing
- LongCat platform docs: model parameters
- LongCat platform docs: model
*Not sure whether a cheap trillion-parameter API beats an open model you can host? Take the free 2-minute quiz and get matched. Then read about the trillion-parameter model that did ship weights this week: Mistral Large 4.*