Are AI token prices subsidised? A cost breakdown for one open model

A Nov 3 newsletter post compares Kimi K3's OpenRouter prices with a self-hosting estimate and argues per-token API prices mostly reflect real costs.

Is per-token AI pricing propped up by venture money? In a Nov 3, 2026 post on the AI for data science newsletter, the author works through the numbers for one open model, Kimi K3. His conclusion, which he calls "my sense," is that per-token API prices mostly reflect real costs. For anyone budgeting agent workloads, the method is the useful part, because you can rerun it for your own model.

He lists 21 OpenRouter providers for Kimi K3 and adds a blended price using agentic-coding ratios: 20:1 input to output, with 95% of input from cache. Blended prices run from $0.59 (Wafer) to $1.81 (AkashML) per million tokens. Six providers, including BaseTen and Moonshot AI, sit at $1.13. He suspects some of the cheapest providers are spending VC cash to win customers.

His self-hosting estimate starts with an eight-GB300 server at about $550,000. Financed, and with colocation, that comes to around $23,000 a month. He takes a throughput figure of about 6,000 tokens/s per GPU from SemiAnalysis rather than measuring it himself. At 75% utilisation, that works out to roughly $0.25 per million tokens, well under $1.13. He used Claude to help, checked individual facts, and warns there may be a deeper flaw. He argues the subsidies sit in subscription plans instead, and notes his analysis ignores training costs. Because Kimi K3 is an open model, it is unclear how far this tells you about closed labs' margins.

The blended view matters. OpenRouter's Oct 28 routing post says its old price-only router let providers list a low input price and recover the difference on cached or output tokens. Buying direct is another option. Per Shattered (Nov 4, 2026), Together AI and Fireworks AI charge no platform fee on serverless usage, while OpenRouter takes 5.5% on credit purchases on its Standard plan.

Sources