Latest in AI
LongCat 2.5 API: Pricing, 1M Context and What Changed
LongCat's new multimodal model combines low introductory API prices with long-running agent workflows, but context capacity and reliable execution are different things.
LongCat announced LongCat 2.5 on November 1, with its API changelog marking availability on November 2. The release targets coding and long-running agent tasks, accepts visual inputs and offers a one-million-token context window. Its introductory API prices make it worth evaluating for workflows that repeatedly read large amounts of context.
LongCat 2.5 API pricing
The official pricing page lists the following rates per million tokens:
| Token type | Limited-time discounted price |
|---|---|
| Uncached input | $0.30 |
| Cached input | $0.006 |
| Output | $1.20 |
These are explicitly discounted prices, not a promise of permanent pricing. The page does not give an expiry date and says the platform and billing records are authoritative.
At those rates, a request using 100,000 uncached input tokens and generating 10,000 output tokens would cost $0.042 in token charges. That is a calculation from the listed rates, not a measured production workload. Retries, additional agent calls and how much context actually qualifies for caching will change the total.
The cached-input rate is particularly relevant when an agent repeatedly works over the same repository or documents. Cheap context helps most when the workflow can reuse it, rather than sending a different large prompt at every step.
Context window, API and visual inputs
The API quick start lists LongCat-2.5 as the model identifier, with a 1M-token context window and a maximum output of 128K tokens. LongCat supports OpenAI-compatible and Anthropic-compatible request formats. That reduces the integration work for an existing client, although compatibility alone does not guarantee identical tool behavior.
The visual input documentation covers images and videos alongside text. Images can be supplied through public URLs or Base64 data; video input uses URLs. Current limits include 10MB per image and 50 images per request, and 100MB per video with up to 10 videos per request.
For a coding workflow, that creates a useful combination: the agent can receive repository context and a screenshot of the interface it is supposed to change. It still needs tools and permissions to inspect files, make edits and run tests. A multimodal model API does not provide that complete execution environment by itself.
What changed in LongCat 2.5
In its release announcement, LongCat describes a mixture-of-experts model with 1.6 trillion total parameters and roughly 48 billion active per token. The company says it expanded training for longer agent trajectories, including turning observed agent failures into tasks that can be checked during training.
LongCat reports a Terminal-Bench v2.1 result of 88.4 for version 2.5, compared with 70.8 for LongCat 2.0. These are vendor-reported results. They support a reason to test the new model, not a conclusion that it will outperform another system in a different coding harness.
The release also describes native four-bit expert computation and inference optimizations. Those implementation choices help explain the company's cost focus, but parameter counts and internal throughput gains are not measurements of the latency an API customer will receive.
The one-million-token context window is not new to this generation: LongCat 2.0 already supported it. The more useful comparison is how reliably version 2.5 uses that context while carrying out a task.
What developers should test first
A large context window makes it possible to provide more material in one request. It does not establish that the model will consistently locate the right detail, follow every constraint or recover from a failed tool call.
For an evaluation, take an existing repository task with a known outcome. Compare completed work, tool failures, retries and total spend, not just the price of the initial prompt. Include a task that requires reading a screenshot if visual coding is part of the intended use.
The interesting question is whether LongCat 2.5's lower input cost survives the full agent loop. If it needs fewer context-management workarounds without adding extra repair attempts, the economic benefit becomes considerably more useful than a cheap token price on its own.
I'm Paras. I work at the intersection of AI, engineering and distribution.
I help technical founders turn model releases, engineering updates and product thinking into clear stories developers want to read and share.