“We, with our agents, hit $300/day per agent using the Claude API, like instantly,” Jason Calacanis said on the All-In Podcast. “And that was doing, maybe, 10 or 20%. That’s $100K/year per agent.” As TradingView/CoinTelegraph reported, Calacanis made the disclosure in February while discussing whether AI agents are actually cost-effective compared to humans at current pricing.

On April 5, Rapid Claw published a breakdown of where that number comes from and what it takes to bring it down.

The Token Math

According to Rapid Claw’s analysis, $300/day requires burning roughly 60 million tokens. At a blended frontier model rate of $5 per million tokens, the progression looks like this:

  • 1M tokens/day → $5/day
  • 10M tokens/day → $50/day
  • 60M tokens/day → $300/day

For a single active agent — not a team — 60 million daily tokens means either very long context windows accumulating over continuous operation, or multiple parallel agents billed as one. Rapid Claw notes that Jensen Huang quantified the agentic multiplier at GTC 2026: agentic tasks consume approximately 1,000x more tokens than a standard chat prompt. For continuous background agents, Huang put that figure at 1,000,000x compared to single-turn prompts. The mechanics: each new completion in a long-running agent includes the full accumulated context. An agent with 200K tokens of working memory pays for those 200K tokens on every subsequent call.

The Routing Opportunity

The core problem, per Rapid Claw, is that most agent deployments send every token to the same frontier model regardless of task complexity. Formatting JSON costs the same as synthesizing six conflicting legal opinions when there is no routing layer.

An estimated 80% of tokens in a typical agentic workflow go to mechanical sub-tasks: parsing responses, reformatting data, classifying results, extracting fields. These tasks do not require a frontier model. A lightweight model at $0.25 per million tokens handles them correctly. Smart routing that dispatches easy tasks to cheap models and reserves frontier-model calls for reasoning-intensive steps brings the $109K/year figure down to an estimated $20,000-$40,000/year — a 60-80% reduction.

Rapid Claw positions its own routing layer as the automated fix. OpenClaw’s built-in multi-model routing serves the same function for operators already on the platform.

What This Means for Builders

Calacanis’s $300/day figure is not an outlier. It is what happens when a high-use agent team runs without cost controls. The number is now circulating publicly and will show up in boardroom conversations about AI infrastructure spend.

For teams planning agent deployments in 2026, the lesson is the same whether you use Rapid Claw, OpenClaw’s routing, or any other dispatch layer: the model selection decision cannot be made once at setup and left alone. Cost scales with token volume, and token volume in continuous agents scales fast.