LongCat 2.5: API pricing, visual inputs and deployment limits
Meituan’s LongCat 2.5 API targets visual coding and longer agent tasks. The launch article is dated November 1, 2026. The API changelog lists availability on November 2. Neither page supplies a timezone.
Meituan describes 1.6 trillion total parameters with about 48 billion activated per token. GitHub and Hugging Face links still say “Coming soon.” Hosted API access does not establish downloadable weights.
Current API prices and account requirements
The pricing page checked November 2 lists these limited-time dollar rates. The page does not identify the dollar currency. It gives no promotion end date and says platform prices and billing records are authoritative.
| Token category | Listed price per million tokens |
|---|---|
| Uncached input | $0.30 |
| Cached input | $0.006 |
| Output | $1.20 |
At those rates, 100,000 uncached input tokens plus 20,000 output tokens cost $0.054. That is an arithmetic example, not a measured agent run. Repeated tool turns can change the total.
The billing guide describes a prepaid balance and requires account verification before topping up. Successful requests are billed by token usage. Failed requests are not charged. Token packs are a separate product, so their quotas should not be confused with the pay-as-you-go balance.
Integration details worth checking
The current quickstart, Responses reference, visual guide and parameter reference describe the following constraints:
| Integration point | Documented behavior |
|---|---|
| Model identifier | LongCat-2.5 |
| API routes | OpenAI-compatible Chat Completions and Responses. Anthropic-compatible Messages |
| Context | 1,048,576 tokens, including input and output |
| Output ceiling | Quickstart says 128K. Parameter reference says 262,144. This discrepancy is unresolved. |
| Images | URL or Base64, up to 50 images, 10 MB each |
| Videos | URL only, up to 10 videos, 100 MB each, each shorter than 3,600 seconds |
| Temperature | Fixed at 1.0 with thinking, 0.7 without. Other values are overridden |
| Responses tool calls | Tool fields are flat, rather than nested inside a function object |
Compatibility labels do not remove migration work. Check the chosen endpoint’s schema, response handling and output limit before moving a production agent. ByteForward has not executed API requests or independently tested the model.
Demos show the intended uses, not independent results
The vendor’s showcase includes a fluid simulation, 3D displays and office workflows. Its benchmark comparisons mix self-run results with figures from external reports and leaderboards, so they should not be read as one controlled test across every model.
Watch the official demonstrations on LongCat’s launch page. These are vendor-provided showcases, not ByteForward performance tests.
For a useful evaluation, start with one existing workflow and its acceptance criteria. Track correctness, retries, elapsed time and billed tokens together. A low token rate alone cannot establish whether an agent finishes a job reliably.
Original AI-generated editorial illustration for ByteForward. Conceptual artwork, not a model output or product interface.
