Astraas — The money layer of AI https://astraas.tech/ AI releases, research and analysis — with the money question asked every time. What it costs, what it replaces, and whether it pays. en hello@astraas.tech (Raunak A.) Mon, 03 Aug 2026 19:22:13 +0000 tools/build-feeds.py The subscription breakeven is 154 million tokens a month. https://astraas.tech/research/subscription-breakeven/ https://astraas.tech/research/subscription-breakeven/ Everyone knows a Claude subscription is cheaper than the API for heavy use. Nobody says where the line is. I measured mine: about 154 million tokens a month. Pricing Raunak A. Tue, 04 Aug 2026 00:00:00 +0000 A cache entry pays for itself on the second read. https://astraas.tech/research/caching-pays-second-read/ https://astraas.tech/research/caching-pays-second-read/ Cache writes cost 25% extra and reads save 90%, putting break-even at 1.39 reads. Across 34 days caching saved $43,327 — 87% of the bill. Prompt caching Raunak A. Tue, 04 Aug 2026 00:00:00 +0000 I routed 7% of my calls to cheap models. It saved 0.89%. https://astraas.tech/research/cheap-models-saved-nothing/ https://astraas.tech/research/cheap-models-saved-nothing/ Model routing is the standard cost advice. Across 1,594 calls, Haiku and Sonnet together came to $57.84 of a $6,465 bill. Model routing Raunak A. Tue, 04 Aug 2026 00:00:00 +0000 My backtest said Sharpe 2.74. Real data said 0.35. https://astraas.tech/research/backtest-said-2-7-reality-said-0-35/ https://astraas.tech/research/backtest-said-2-7-reality-said-0-35/ Nineteen systematic strategies tested. Fourteen died to transaction costs. The one that looked spectacular was a proxy — and rebuilding it on real option prices cut the number by 87%. Cost floors Raunak A. Tue, 04 Aug 2026 00:00:00 +0000 Inference got 95% cheaper. Your bill went up anyway. https://astraas.tech/research/prices-fell-bills-rose/ https://astraas.tech/research/prices-fell-bills-rose/ GPT-4-class inference fell from $30 per million tokens to under $0.50 in two years. Enterprise spend on it roughly doubled. Agentic workloads explain the gap. Inference economics Raunak A. Tue, 04 Aug 2026 00:00:00 +0000 95% of enterprise AI pilots return nothing. The failure isn't the model. https://astraas.tech/research/95-percent-return-nothing/ https://astraas.tech/research/95-percent-return-nothing/ MIT studied 300 enterprise AI deployments and found 95% delivered no measurable P&L impact. What the surviving 5% did differently has little to do with model choice. AI ROI Raunak A. Tue, 04 Aug 2026 00:00:00 +0000 I logged 9.96 billion tokens in 34 days. Output was 8% of the bill. https://astraas.tech/research/token-bill/ https://astraas.tech/research/token-bill/ Every guide to cutting LLM costs opens with the same advice: shorten your outputs. I costed 34 days of real agentic work line by line. Output tokens were 8% of spend. Cache reads were 75%. Agent economics Raunak A. Mon, 03 Aug 2026 00:00:00 +0000