Astraas·Section

AI

Models, tools, and the money behind them. What shipped, what it costs, and whether it actually works yet.

• 4 Aug The subscription breakeven is 154 million tokens a monthPricing Everyone knows a Claude subscription is cheaper than the API for heavy use. Nobody says where the line is. I measured…
• 4 Aug A cache entry pays for itself on the second readPrompt caching Cache writes cost 25% extra and reads save 90%, putting break-even at 1.39 reads. Across 34 days caching saved $43,32…
• 4 Aug I routed 7% of my calls to cheap models. It saved 0.89%Model routing Model routing is the standard cost advice. Across 1,594 calls, Haiku and Sonnet together came to $57.84 of a $6,465 b…
• 4 Aug My backtest said Sharpe 2.74. Real data said 0.35Cost floors Nineteen systematic strategies tested. Fourteen died to transaction costs. The one that looked spectacular was a prox…
• 4 Aug Inference got 95% cheaper. Your bill went up anywayInference economics GPT-4-class inference fell from $30 per million tokens to under $0.50 in two years. Enterprise spend on it roughly do…
• 4 Aug 95% of enterprise AI pilots return nothing. The failure isn't the modelAI ROI MIT studied 300 enterprise AI deployments and found 95% delivered no measurable P&L impact. What the surviving 5% did…
• 3 Aug I logged 9.96 billion tokens in 34 days. Output was 8% of the billAgent economics Every guide to cutting LLM costs opens with the same advice: shorten your outputs. I costed 34 days of real agentic w…