LongCat-2.0 Official API at $0.75/$2.95 + SiliconFlow Parity: Where's the Self-Host Break-Even?

Two Prices, One Model, One Big Decision

Meituan's LongCat-2.0 — the 1.6T-parameter (48B active) agentic coding MoE released June 30, 2026 — now has two clear inference paths priced within a few percent of each other:

• Meituan official API: $0.75 input / $2.95 output per M tokens (regular). Cached input $0.015/M. Promotional pricing $0.30/$1.20 for a limited window.
• SiliconFlow hosted: ¥5 input / ¥20 output per M tokens ≈ $0.70/$2.82 at current CNY/USD.
• Self-host: Weights are open (BF16/FP8/INT8) under a permissive license on Hugging Face at meituan-longcat/LongCat-2.0.

For most teams the question is not "which API is cheapest" — the API prices are within 5% of each other. The real decision is when does self-hosting the weights pay off versus paying $0.75/$2.95 per million tokens?

The Hardware Reality

LongCat-2.0 is 1.6T total parameters with ~48B active per token. The BF16 weights are approximately 3.2 TB on disk; the FP8 quantized version is about 1.6 TB; INT8 lands around 800 GB. In practical terms:

• FP8 inference server: 8× H100 or 4× B200 as the minimum viable production node. Hardware capex $200,000-$400,000 or ~$8,000-$15,000/month on cloud (AWS p5, Azure ND-H100-v5, or GCP a3-highgpu).
• INT8 inference: Squeezes onto 4× H100 or a single B200 8-GPU node with reduced batch size. Cloud cost $4,000-$8,000/month for a single-node setup.
• Domestic Chinese hardware (Ascend, Cambricon): Meituan trained and validated on 50K+ domestic AI ASIC superpods, so INT8 inference on Chinese silicon is production-ready. Cost roughly ¥30,000-¥60,000/month per node.

The Break-Even Math

Assume you have a cloud FP8 node costing $10,000/month all-in (GPU, storage, networking, ops overhead). At $0.75 input / $2.95 output, how many tokens must you push through it to break even against the API?

Assuming a typical agentic coding traffic mix of 4:1 input-to-output tokens:

• Blended API cost = (4 × $0.75 + 1 × $2.95) / 5 = $1.19 per M blended tokens
• Monthly break-even token volume = $10,000 / $1.19 = ~8.4 billion tokens/month

8.4 billion tokens/month is a lot: roughly 280 million tokens/day, or about 5,000-8,000 substantial agentic coding tasks per day depending on your average task size. For most teams that's 100+ full-time developers running Claude Code or Cursor continuously. Below that scale, the API wins.

Why the Promo Price Matters (and Why It Doesn't)

The promotional $0.30/$1.20 pricing cuts the break-even point roughly in half — to about 17 billion tokens/month, which is even higher volume than the regular price. But promotional pricing is not a planning input for capex decisions. The disclaimer on Meituan's pricing page explicitly says prices are subject to change. Build your break-even model against the $0.75/$2.95 regular price and treat the promo as a temporary bonus.

The Non-Cost Reasons to Self-Host Anyway

Break-even math ignores the non-cost variables that push teams toward self-hosting even when they lose the token arithmetic:

• Data residency: Financial services, healthcare, and Chinese enterprise customers with domestic-only data mandates cannot send code to a Meituan-managed API even if the price is right. Self-host wins by default.
• Latency-sensitive workflows: A dedicated on-prem or same-region inference server can hit 100-200ms first-token latency versus 400-800ms for a shared API. For interactive coding, that matters.
• Fine-tuning access: Self-hosting lets you SFT or continue-train on your codebase. The API doesn't offer that today.
• Predictable spend: A fixed $10,000/month bill is easier to budget than variable API charges that spike with usage. This matters for cost predictability even when the average price is higher.

The Cache Story

LongCat-2.0's cached input at $0.015/M is 50x cheaper than uncached input. For agentic coding workloads with 85%+ cache hit rates, the effective blended input price drops from $0.75 to roughly $0.13/M. That pushes the API break-even for self-hosting even higher — to about 15-20 billion tokens/month at typical cache rates.

Self-hosting also has caching (via SGLang or vLLM's prefix cache), but the operational overhead of keeping cache warm across restarts and node failures is a real cost. The API abstracts that away.

Practical Recommendation

For teams under ~100 active AI-coding developers or under ~5 billion tokens/month: stay on the Meituan or SiliconFlow API. The break-even math doesn't work out and you avoid ops complexity.

For teams with data-residency constraints, sensitive IP, or 200+ developers running agents continuously: start planning a self-host. Begin with a 4-week pilot on INT8 to validate quality, then scale to FP8 in production. Budget $150-$300K in year-one capex or $10-15K/month cloud opex plus 0.5-1.0 FTE of ML platform engineering.

Everyone else: enjoy the price war. Two hosted providers competing for your dollars at parity is the best possible market structure for buyers.