ModelBrief · LongCat 2.5 · Topic Page
Last updated Nov 13, 2026

LongCat 2.5 Pricing: $0.30/$1.20 Promo and the Catch

ModelBrief · Published 2026-11-13 · Fact-check 2026-11-13 · Reference: ToolPilotPro pricing guide

LongCat 2.5's pricing went live with the API platform's GA on November 2, 2026: $0.30 in, $0.006 cached in, $1.20 out per million tokens (¥2 / ¥0.04 / ¥8 in RMB terms). All three figures are marked as limited-time discounted prices — but there is no end date, no list price, and the pricing page only notes that prices may change. The same-priced reference point is DeepSeek V4.1 Flash, whose peak pricing matches these three tiers exactly and drops to half price off-peak. The cheapness is real; how long it lasts is unknown — that is the premise for reading this price list.

Three tiers, and why the cache price matters

The three tiers are: uncached input $0.30 (¥2), cached input $0.006 (¥0.04), output $1.20 (¥8). The cached price is just 2% of the uncached price — one-fiftieth. For long-horizon agents that repeatedly read the same repository, that ratio matters: once the large repeated chunks of context hit the cache, input cost becomes almost negligible — the more a model is built for long tasks, the more the cache discount is worth in practice. Note that all three tiers are limited-time discounts: the pricing page gives no end date and no post-discount list price, only that prices may change.

Billing rules: pay only on success

Several billing rules work in the user's favour: only requests that return HTTP 200 are billed — failed requests cost nothing; prepaid balances never expire; promotional credit is used before paid credit; there is no surcharge tied to context length; and image and video inputs carry no separate charge — they are billed only as the tokens they convert into. The official worked example: 600K uncached input + 400K cached input + 500K output comes to $0.7824 — under a dollar for a mid-sized agent task at that mix.

Volume discounts, Token Pack and payment

Verified companies get volume discounts on monthly spend: 10% off above $75 (¥500), 20% off above $300 (¥2,000), 30% off above $1,200 (¥8,000), 40% off above $6,000 (¥40,000). There is also a Token Pack: a fixed 30-day quota sold in flash-sale windows at 10:00, 16:00, 21:00 and 23:00 Beijing time; sizes and prices are only visible inside the console after login and are not public. On payment, the USD site accepts credit cards, debit cards and Google Pay (the 2.0 era supported only Alipay and WeChat Pay); outside mainland China you can register with an email address, with no identity verification.

The repricing risk: no end date, no list price

The biggest uncertainty is what happens after the discount. The old 2.0 list price was $0.75 / $0.015 / $2.95 — if 2.5 moves to the same level when the discount ends, the output price would rise about 2.5×. But that is speculation: Meituan has never said what 2.5 will cost after the discount, and 2.5 has no list price of its own. For comparison, DeepSeek V4.1 Flash matches LongCat 2.5 exactly at peak and halves off-peak, while Qwen3.8 Max ($2 / $6) and Kimi K3.1 Pro ($2 / $10) cost several times more. A community post claims new users get a 10M-token starter credit and existing users 5M, with a ¥9.9 limited-time offer coming soon — that claim has no official page behind it, is unconfirmed, and should not be used for planning.

ModelBrief's take: At the current limited-time price, LongCat 2.5 is one of the cheapest long-horizon agent models on the market, and a cache price at 2% is a structure built for long tasks. But the price list has two blanks: when the discount ends, and what comes after. Budget-sensitive teams can start using it now — and should run one stress test at the 2.0 list price: if an output price of $2.95 is still acceptable, the discount is safe to rely on.

Reference: ToolPilotPro — LongCat-2.5 Guide (rewritten and organized by ModelBrief; facts per the official source)