A March 9 Hacker News discussion centered on Martin Alderson's post pushing back on a viral claim that Anthropic is effectively spending about $5,000 in compute for every Claude Code Max subscriber paying $200 per month. The core argument is not that heavy users are cheap to serve, but that the headline number likely mixes up retail API-equivalent usage with Anthropic's own underlying inference cost.

Alderson's breakdown starts from Anthropic's public Opus 4.6 pricing: $5 per million input tokens and $25 per million output tokens. At those rates, he argues, an extremely heavy Claude Code user really could burn through something in the $5,000 range in API-equivalent value. His key point is that list price is a billing layer, not necessarily a proxy for what it costs Anthropic to run the model internally.

To make that distinction concrete, the post compares Opus pricing with OpenRouter listings for large open-weight MoE models. It cites Qwen 3.5 397B-A17B at $0.39 per million input tokens and $2.34 per million output tokens, plus Kimi K2.5 at $0.45 and $2.25 respectively. From that comparison, the author infers that the actual serving cost for the heaviest Claude Code users could be closer to hundreds of dollars per month than thousands, especially if cache reads and average utilization are taken into account.

The post also points to Anthropic's own /cost command data, which it says puts average Claude Code usage around $6 per day in API-equivalent spend, with 90% of users under $12 per day. Under the assumptions in the essay, that would make average subscribers much closer to break-even or profitable, while only a narrow slice of power users would be meaningfully subsidized.

What made the HN thread valuable is that readers immediately stress-tested those assumptions. Some argued that Chinese open-weight models are not a fair comparison because infrastructure, model size, and inference efficiency can differ materially. Others pointed to Bedrock and Vertex throughput as evidence that Opus is unlikely to be an order of magnitude more expensive to serve than leading open-weight peers. The result was less a consensus than a useful reframing of the debate.

For developers building on frontier APIs, that reframing matters. A tool vendor paying near-retail API rates can have painful unit economics even if the model provider itself is not losing money on average usage. The HN discussion therefore landed on a more precise takeaway: API-equivalent spend, provider serving cost, and subscription economics are different layers, and product decisions around Claude Code, Cursor, or open-weight alternatives only make sense when those layers are separated explicitly.

#pricing

HN Debates Whether Claude Code's '$5k User' Meme Confuses API Pricing With Real Inference Cost