# AI and cloud cost research > Canonical: https://bytecosts.com/research/ **Direct answer.** This is the ByteCosts research index: 17 interactive data stories on what AI and cloud actually cost, written from the committed ByteCosts pricing dataset. Each deep-dive pairs an interactive scrollytelling article with a crawlable summary page, and every quantitative claim is source-backed or labeled as modeled. Use this page to find the cost pattern you are about to budget for - per-user unit economics, subscription break-evens, caching, GPU rent-vs-own, platform bills - then open the matching calculator and replace the example assumptions with your own workload. The full list, grouped by topic, is in the static HTML before JavaScript runs. ## What this index is ByteCosts publishes 17 research deep-dives built from the same committed pricing data that powers the calculators. Each piece below links to a summary page with the article's method, sources, and citation block, plus the interactive data story. Every article follows the same honesty rule: numbers are derived from committed source-backed rates, externally cited, or explicitly labeled as modeled scenarios. ## How to use the research - Pick the cost pattern closest to the decision you are making - Read the summary page first; open the interactive story for the full walkthrough - Bring the pattern back to a ByteCosts calculator and swap in your own usage - Cite the summary page, not the interactive file - it carries the canonical URL ## Frequently asked questions ### How is this different from the ByteCosts blog? The research index lists the interactive data-story deep-dives. The blog at /blog carries editorial posts and analysis written around the same dataset. ### Are the numbers in these articles real prices? Each article states its basis: figures are committed-data-derived, externally cited, or labeled as modeled. The per-article summary page lists sources and method. ## Unit economics - [AI cost per active user](https://bytecosts.com/research/ai-cost-per-active-user/) - A token price is not a business metric. Scroll one AI app from raw inference to real per-user COGS and watch - [The margin bridge](https://bytecosts.com/research/the-margin-bridge/) - Take $100 of AI-product revenue and walk it down to what you actually keep. Each cost layer is a step on the ## AI economics - [Anatomy of an AI invoice](https://bytecosts.com/research/anatomy-of-an-ai-invoice/) - Everyone budgets for inference. Then the bill arrives and the model tokens are the small slice. Watch a - [The real cost of AI coding assistants](https://bytecosts.com/research/ai-coding-assistants-cost/) - A $20 seat is the headline. Across the five major assistants’ published seat prices plus a modeled usage ## SaaS economics - [How a $99 tool becomes a $420 tool](https://bytecosts.com/research/hidden-saas-costs/) - The number on the pricing page is an opening bid. Seat minimums, usage overages, and the enterprise-feature ## Macro - [LLMflation](https://bytecosts.com/research/llmflation/) - The cost of a token has fallen perhaps a thousand-fold in three years, one of the steepest price declines in ## AI infrastructure - [Rent, reserve, or own](https://bytecosts.com/research/rent-reserve-or-own/) - The cheapest way to buy a GPU-hour depends entirely on one number: how much of the time the GPU is actually ## Orchestration - [The 15× Multiplier](https://bytecosts.com/research/the-15x-multiplier/) - Hand a hard research question to a team of agents instead of one and the answer gets ninety per cent better ## Cost dynamics - [The 2.7× you didn’t budget for](https://bytecosts.com/research/cost-story/) - An AI feature’s sticker price is the cheapest number you’ll ever see it post. Follow one feature from ## Total cost of ownership - [The 25 Per Cent Bill](https://bytecosts.com/research/the-25-percent-bill/) - Finance approves the build. The build is barely a quarter of what a production agent costs over three years ## Market economics - [The great model price collapse](https://bytecosts.com/research/model-price-collapse/) - Frontier inference is getting roughly 10× cheaper every year. Scroll the last two years and watch the ## Infra economics - [The observability bill](https://bytecosts.com/research/the-observability-bill/) - Monitoring starts as a rounding error and ends as one of your largest infra line items. Scroll your fleet ## Engineering - [The Optimization Playbook](https://bytecosts.com/research/the-optimization-playbook/) - There are six levers that reliably cut an agent’s bill, and they are not equally powerful. Pull them in the ## Token mechanics - [The Quadratic Trap](https://bytecosts.com/research/the-quadratic-trap/) - A ReAct agent re-sends everything it has ever seen on every call. That one fact turns a ten-step task into a ## AI architecture - [The RAG tax](https://bytecosts.com/research/the-rag-tax/) - “Just add retrieval” sounds free. In practice, grounding a model in your own data wraps the inference quote ## Cloud economics - [Vercel vs AWS vs Railway, on published rates](https://bytecosts.com/research/cloud-cost-showdown/) - One production app, three platforms, priced on each platform’s published rates across a modeled growth curve ## Benchmarks - [What a Task Actually Costs](https://bytecosts.com/research/what-a-task-costs/) - Price six recognizable 2026 flagships on a single coding task and the cost spreads nearly seventeenfold ## Cite this page ByteCosts research index. ByteCosts. https://bytecosts.com/research/ **Sources** - [ByteCosts methodology](https://bytecosts.com/methodology/) - [ByteCosts blog](https://bytecosts.com/blog/)