LongCat-2.5 Review: 1.6T Parameter Multimodal Model for Long-Horizon Agent Tasks

LongCat-2.5 launched as generally available on the LongCat API Platform on November 2, 2026. The model carries 1.6 trillion total parameters with roughly 48 billion active per token and targets long-horizon agent tasks that combine text, images, and tool calls over extended sessions.

Model Architecture, Context Window, and Core Capabilities

Meituan developed LongCat-2.5 as a native multimodal mixture-of-experts model. It processes text and image inputs through discrete visual tokens generated by multi-level residual vector quantization over native-resolution features. The unified backbone also accepts video in some configurations and produces text output only. A dedicated thinking mode and tool-use support allow the model to handle sequential decision steps typical of agent workflows.

The native context window reaches one million tokens, while maximum output length ranges from 128,000 to 256,000 tokens depending on the endpoint configuration. Training incorporated synthetic trajectories drawn from GitHub pull requests and issues, together with pass-to-pass verification across multiple harnesses. These additions build directly on LongCat-2.0, released with open weights in June 2026, and add native multimodality plus scaled reinforcement learning for longer agent episodes. Infrastructure changes include FSDP and long-context optimizations run on domestic Chinese accelerators.

Vendor benchmarks show clear gains over the prior version. On Terminal Bench v2.1 the score rose from 70.8 to 88.4; on v4.0 it moved from 3.0 to 38.5. DeepSWE v1.1 improved from 9.7 to 64.6, BrowseComp from 79.9 to 91.9, and NL2Repo from 40.9 to 60.5. The model sits near Claude Opus 5 and DeepSeek V4.1 Flash on several agent and browsing tasks but trails on harder coding subsets. W4A4 quantization supports higher inference throughput.

API Access, Pricing Structure, and Performance Comparison

The LongCat API accepts requests in both OpenAI-compatible /v1/chat/completions and Anthropic-compatible /v1/messages formats at the endpoint https://api.longcat.chat. The model identifier is LongCat-2.5. Integrations exist with agent frameworks such as Claude Code, Codex, and Hermes. Rate limits trigger HTTP 429 responses when exceeded; documentation suggests a retry interval near 60 seconds, though exact requests-per-minute and tokens-per-minute figures remain unpublished.

Pricing follows a pay-as-you-go schedule with a current limited-time discount. Input tokens cost $0.30 per million uncached or $0.006 per million cached. Output tokens cost $1.20 per million. Billing occurs strictly on actual usage; no monthly subscriptions or fixed plans appear in the published schedule. The discount period has no announced end date.

Independent third-party benchmarks remain scarce as of early November 2026. All published scores originate from Meituan internal evaluations. Open weights for LongCat-2.5 have not been released, though the announcement states they will appear on GitHub and Hugging Face at a later date. LongCat-2.0 continues to serve as the open flagship model.

Current Limitations and Practical Considerations for Developers

Public documentation does not detail security certifications, data residency options, or third-party audits. Users must rely on Meituan’s stated infrastructure practices until independent reports appear. Rate-limit behavior and retry logic require testing inside production agent loops, especially for workloads that exceed typical token budgets.

Multimodal input currently routes through the API as discrete visual tokens; developers must prepare images or video frames accordingly. No fixed pricing tiers exist, so costs scale directly with token volume. The absence of published open weights limits local experimentation and fine-tuning options at present.

LongCat-2.5 suits developers who need extended context and native multimodality for agent tasks and who can accommodate vendor-reported metrics until broader benchmarks emerge. The pay-as-you-go model keeps entry costs low for variable workloads, yet output pricing remains higher than several competing text-only APIs. Continued monitoring of weight releases and independent evaluations will clarify long-term viability.