LongCat-2.0 API Is Now on SiliconFlow: The Agentic Coding Routing Decision for 1M-Context Workloads

For weeks, engineering teams using Hermes Agent and Claude Code were trusting an anonymous model on community leaderboards without knowing its name. "Owl Alpha" — the mystery model that reached Top 3 globally by daily volume, ranked #1 on Hermes Agent and #2 on Claude Code by monthly usage — was a preview of Meituan's LongCat-2.0. Now the full model is public, the API is open via SiliconFlow, and every operator needs to make a deliberate decision about where it fits in their routing chain.

What happened

Meituan's LongCat team released LongCat-2.0 on June 30, 2026, followed by its availability on SiliconFlow as an OpenAI- and Anthropic-compatible API on July 5, 2026 (with a blog update July 17). The model is:

• 1.6T total parameters, ~48B active per token — a sparse MoE architecture with a dynamic activation range of 33B–56B
• Native 1M-token context window (1,049K tokens exactly), with 262K max output
• LongCat Sparse Attention (LSA): linear-scale compute for long-context, not quadratic — precision doesn't degrade at the full 1M window
• MOPD routing: three expert groups (Agent / Reasoning / Interaction) independently trained and dynamically gate-scheduled at inference
• SWE-bench Pro leader — top performance on long-horizon software engineering benchmarks

Pricing on SiliconFlow:

• Input: $0.75/M tokens
• Cache read: $0.015/M tokens (50× cheaper than uncached input)
• Output: $2.95/M tokens

Why it matters for AI engineering teams

LongCat-2.0 occupies a specific position in the model market: it is not the cheapest option (GLM-5.2 at ~$0.20/M input is considerably cheaper) and not the fastest (MiMo-V2.5-Pro-UltraSpeed targets real-time throughput). What LongCat-2.0 delivers is the combination of 1M-context reliability + agentic coding quality that its anonymous leaderboard run actually validated in production.

The SWE-bench Pro result matters because that benchmark is harder to game than SWE-bench Verified — it tests entire software engineering sessions, not single file edits. If your workload is a coding agent that needs to navigate a large codebase, write tests, and verify results across multiple tool calls, this is a direct signal.

The cache pricing is the key operator variable. At $0.015/M for cache hits, a Claude Code session that keeps the codebase in the prefix cache will pay far less than the headline input price suggests. For agents running repeated calls against the same 300K-token context, the effective cost drops significantly.

The router/operator angle

Where LongCat-2.0 fits in a routing policy:

1. Primary coding-agent tier — for teams whose primary signal is SWE-bench Pro performance and whose sessions exceed 200K tokens. At this context length, models that can't handle it reliably fail; LongCat-2.0's LSA architecture is specifically designed for this.

2. Fallback for GLM-5.2 long-context failures — GLM-5.2 is faster and cheaper, but its 1M context is newer; LongCat-2.0's architecture was purpose-built for long-context from the ground up. Use LongCat-2.0 as the fallback route when sessions hit the GLM-5.2 context edge.

3. Anthropic-compatible API drop-in — SiliconFlow exposes both OpenAI-compatible and Anthropic-compatible endpoints. This means no SDK changes for Claude Code users pointing at a custom base URL.

Cost comparison for a 1M-context session (no cache hits):

• LongCat-2.0 on SiliconFlow: ~$0.75 input
• GLM-5.2 on DashScope: ~$0.14 input
• DeepSeek-V4-Pro: ~$0.27 input

LongCat-2.0 is approximately 3–5× more expensive than the cheapest Tier 2 alternatives on uncached input. Whether that premium is justified depends on whether your workload actually needs SWE-bench Pro quality on long-horizon coding. For most standard code completion, it won't be. For autonomous agent sessions running across a full monorepo, it might.

What to watch:

• Open weights July 27 — Meituan has committed to releasing the full model weights by July 27, 2026. This opens up self-hosted deployment paths, VLLM/SGLang integrations, and theoretically cheaper inference at scale.
• Context cache stability — The $0.015/M cache rate is currently a SiliconFlow-set price. If open weights allow operators to run LongCat-2.0 on their own infrastructure, the economics shift dramatically.
• Tool-use reliability at depth — SWE-bench Pro measures structured coding sessions. Operators should eval their specific tool-use schema against LongCat-2.0 before routing production traffic.

What TheRouter users should watch or try

If you are managing routing across multiple providers, LongCat-2.0 via SiliconFlow is a new OpenAI-compatible provider endpoint worth adding as a candidate for high-context coding agent routes. The model ID on SiliconFlow is meituan-longcat/LongCat-2.0.

Add it to your routing table with a context-length guard: route requests with context_tokens > 200000 to LongCat-2.0 when the task type is coding/agent, and fall back to GLM-5.2 or DeepSeek-V4-Pro for shorter sessions where the cost premium isn't warranted.

See the providers documentation for configuring a custom OpenAI-compatible provider endpoint and context-window routing guards.