What It Cost to Build Four Free SEO Tools With Claude, and the $73 Retry

Building four free SEO tools with Claude took between 112 million and 170 million tokens a tool, which comes to about $34 to $50 each at Anthropic’s list prices for Opus 5.5. Those are list-price equivalents worked out from the session logs, not invoices. The run that cost more than any of the builds wasn’t a tool at all. It was a 20-minute retry of a site redesign that spent about $73 and left nothing I kept.

The tools are part of GEO Search Console, my free tool site for AI search. Most of the ideas behind its tools are David Quaid’s, and some are Jonathan Denwood’s. The building was mine, with help from Claude.

What each free SEO tool build cost

What each free SEO tool build cost came down to tokens, and nearly all of them were the same kind. Each tool was built in one Claude Code session. On October 11 I added up the usage in those four sessions’ logs, counting each message once.

Tool	Million tokens
Knowledge Panel Checker	134
AI Performance Report	120
AI Citation Gap Tool	112
AI Brand Fact Checker	170

Output was small: 0.31 million to 0.44 million tokens a build. One-hour cache writes ran from 0.70 million to 0.81 million. Nearly all the rest, about 99 percent of the total, was cache reads.

Why rereading context is most of the bill

Why rereading context is most of the bill is plain arithmetic. Each step in a session rereads the conversation so far, and in these sessions every step reread about 480,000 to 590,000 tokens of context. Over a few hundred steps, that’s where a hundred million tokens come from.

Cache reads are the cheapest tokens on Opus 5.5’s price list: $0.20 a million, against $4 for fresh input, $8 for one-hour cache writes and $20 for output. At that volume they still carry the bill. For the biggest build, roughly 168 million cache reads come to about $34 before a single output token is counted.

These sessions were also heavier than they had to be. Each one was forked from a long stretch of earlier work, so it carried a lot of that earlier context into every step. A fresh, lean session rereads less on every step, so it would most likely cost less for the same tool.

The $73 retry

The $73 retry happened on October 11, on a redesign of GEO Search Console’s look rather than a tool. It was a multi-agent workflow: seven families of pages, each with a build, two reviews and a fix. When my account hit its usage limit partway through, 9 of the 25 agents failed.

I said try again, and Claude resumed the run. A resume reuses only the longest unchanged run of finished calls from the start, in call order. The first failed call was the third one, so everything from there on started over. Of the nine agents the resume started, seven were redoing work that had already finished.

Seven of the nine agents were redoing work that had already finished.

In about 20 minutes it spent about $73. Claude stopped it and told me plainly what had happened. Nothing from that run was kept.

Why the estimate missed

Why the estimate missed is the part I’d tell anyone who runs agents in parallel. Claude’s estimate to finish had been $25 to $45. It came from token counts with a cheap cache-read price assumed, not from a spend rate anyone had measured. The measured rate turned out to be about $3.50 a minute with nine Opus agents running at once, or about $210 an hour.

For scale, another multi-agent workflow on the same redesign made 1,465 model calls and would have come to about $165 at list prices: 472 million cache reads, 8.6 million five-minute cache writes and 1.4 million output tokens. My read is that parallel agents multiply the rereading, since each one carries its own context.

What I do differently now

What I do differently now starts before anything runs.

• After a partial failure, I write a new run that does only the missing pieces, worked out from the record of what finished. I don’t resume when the first failure sits early in call order.
• Before trusting a cost estimate, I measure spend at the start and again a few minutes in, and go by that rate.
• If a run would go past my plan’s limit, the default is to wait for the reset.

Starting a new tool in a lean session is most likely cheaper than forking a long one. Running the tools is a separate bill: on launch day, two of the site’s other tools used up a $5 daily cap by about 9 AM.

How I measured the builds

How I measured the builds was by reading the logs Claude Code keeps for every session.

To run the same count on your own work, sum the usage on every message in a session’s log, including any subagents’ logs, and keep one entry per message ID, since the same message can be logged more than once. Then price each kind of token separately. A total token count on its own hides the fact that a million output tokens cost a hundred times what a million cache reads do.

The tools themselves are on the My Tools page, with what each one does.

Sources

1. Pricing (Claude Platform Docs)
2. GEO Search Console (GEO Search Console)