Token prices keep falling. Search prices mostly haven't. Enter your model, token volume and search vendor to see how much of your agent bill is search today, and how that share moves over the next three years.
The selected vendor drives the stacked chart. List prices, editable.
Pick up to five vendors. Same workload, growth and search price trend for each.
Search costs the same whichever model runs the agent, so the cheaper the model, the more of the bill is search. Top row uses your prices above; the rest use list prices.
How the math works. Inference = tokens × blended price (input share × input price + output share × output price). Search = tokens × searches per 1M tokens × vendor price per 1K. Each month, token volume grows and token price falls at the annual rates you set; search volume grows with tokens and its unit price moves at the search rate.
Where the defaults come from. Model prices are standard-tier list prices as of 11 Nov 2026, with no caching, batch or long-context adjustments; OpenAI, Anthropic, Google and DeepSeek were checked against their own pricing pages, the rest come from a third-party price tracker. Gemini 4 Argon and 3.8 Flash are on promotional rates that double on 1 Dec 2026 and 1 Jan 2027; DeepSeek is shown at peak rates. Search prices are each vendor's published pay-as-you-go rate for a basic search of up to 10 results (Parallel, Exa, Tavily, Brave, Perplexity, Linkup); Firecrawl is derived from its plan credits. The deep research preset uses Google's published estimate for its Deep Research agent, about 80 searches per roughly 310K tokens. The other presets are rough assumptions. The 3× yearly token price drop is deliberately slower than the 10× per year a16z measured at constant capability, since teams tend to upgrade models. Everything is editable; verify prices before quoting them.