AEOSearch

Blog · 2026-08-21 · 6 min read

Models charge what they want

We pay retail for every grounded answer, across five AI engines, and read the bill line by line. The lesson: a model's per-token price tells you almost nothing, because the model decides how much of your money to spend on each question.

Technical information first: we have run thousands of grounded answers through five AI engines — ChatGPT, Gemini, Claude, Grok and Perplexity — so we see the differences in pricing and performance from the inside. The interesting conclusion we can draw is that each model chooses how much to spend on each question, often arbitrarily so.

AEOSearch audits ask the answer engines your buyers actually use tangible questions with live web search on, and we pay retail pricing for every call so that we emulate the buyer experience as closely as possible. When we probe our own costs we read back exactly what each provider reports — input tokens, output tokens and search calls — and the raw runs are logged, so we can calibrate our work to produce the most realistic results on every subsequent audit.

We have learned that for a grounded question, the per-token price of a model is nearly entirely disconnected from the cost of the answer. The only thing that matters is what the model chooses to do to answer our query. The vendors say as much on their own price lists. OpenAI bills web search at "$10.00 / 1k calls + Search content tokens billed at model rates" (OpenAI API pricing); Anthropic charges "$10 per 1,000 searches, plus standard token costs for search-generated content" (Anthropic, Web search tool). On a frontier model the search fee is the cheap part. What the model drags back in, and how many times it goes, is the bill.

Our three recent surprises

Inquiries with the same model for the same task saw a 1.8x spread. We probed OpenAI's frontier model (gpt-5.6-sol) on four different grounded questions. For each of the 4 questions OpenAI chose to issue 3, 4, 5 and 6 web searches, so the four answers cost $0.208, $0.253, $0.311 and $0.371, respectively. Cost was monotone in the search count. There is no single "cost per answer" for that model; there's a band, and the model picks where in the band your question lands. Meanwhile gpt-4o-mini issued exactly one search on every probe and came in at about $0.0116 per answer, proving the stark difference in consumer model behavior. Keep in mind that we use frontier models for the $199 audit and every subscription so that we emulate the behavior of your consumers — the report names which model answered, per engine — but we can also tailor an audit to your needs.

Even more strangely, one model in a family can be an order of magnitude off its siblings. A grounded claude-sonnet-5 answer billed 109k–138k input tokens: the engine re-sends accumulated context each search round, and it compounds. Anthropic's own documentation spells out the mechanism: "Web search results retrieved throughout a conversation are counted as input tokens, in search iterations executed during a single turn and in subsequent conversation turns" (Anthropic, Web search tool). Claude Haiku 4.5, same job, used 9.2k–10k. That's an 11–14x cost difference inside one vendor, on one web search either way. It's the specific reason our free tier doesn't run Anthropic at all: Sonnet 5's grounded-answer cost doesn't fit a $0 report, and Claude coverage answering on the frontier model is part of what the paid tiers buy.

Most bizarrely, the older model is the expensive one. Google prices grounding by model family: on its price list Gemini 2.5 Flash pays "$35 / 1,000 grounded prompts" and the Gemini 3.x models pay "$14 per 1,000 requests" (Google, Gemini Developer API pricing). Measured per answer, gemini-2.5-flash cost $0.0365–$0.0368 against gemini-3.6-flash's $0.0211–$0.0217. "Downgrade the free tier to the older model" would have paid roughly 70% more for a staler measurement — so every tier, free included, answers on 3.6 Flash.

Why we publish this

Because it is important for consumers to understand how quickly this technology is changing, and it is our job to stay ahead of the trends to solve your problems as efficiently as possible. Stanford's 2025 AI Index reports that "the inference cost for a system performing at the level of GPT-3.5 dropped over 280-fold between November 2022 and October 2024" (Stanford HAI, The 2025 AI Index Report). And yet a grounded answer from a frontier model still costs real money in 2026, because the model is doing more work per question, not less: the search calls, the retrieved pages, the re-sent context.

This is the honest answer to "why does the $199 audit cost $199?" A full paid audit is 180 grounded answer units across three engines, and the two dominant cost terms are bands driven by model behavior we don't control. And it is still the best value in the product category today: that $199 buys 180 frontier-model answers, every one stored and one shown verbatim per question, plus a 90-day action plan built from the questions you lost. We provide you with incalculable value for that $199, and we solve your problems directly if you choose to engage us to do so.

We are willing to use this technology on our own products to prove our capabilities, and we are in the process of doing so. We look forward to growing with you, and our agents are at your beck and call. Start with the free scorecard, or read why AEO is eating SEO first.

Sources

  1. 1. OpenAI, "Pricing," OpenAI API documentation (web search tool: "$10.00 / 1k calls + Search content tokens billed at model rates"), accessed August 22, 2026
  2. 2. Anthropic, "Web search tool," Claude Developer Platform documentation ("$10 per 1,000 searches"; search results "are counted as input tokens"), accessed August 22, 2026
  3. 3. Google, "Gemini Developer API pricing" (Grounding with Google Search: Gemini 2.5 Flash "$35 / 1,000 grounded prompts"; Gemini 3.x "$14 per 1,000 requests"), accessed August 22, 2026
  4. 4. Stanford Institute for Human-Centered Artificial Intelligence, "The 2025 AI Index Report," Takeaway 7: AI becomes more efficient, affordable and accessible

— The AEOSearch team

Start with the free scorecard.

Your buyers are already asking the engines. Get your grounded baseline.

$0 to start · no card required · every number traces to a stored answer