A new computer use agent cost claim is making the rounds: Anchor, a New York company that calls itself “the browser infrastructure company for AI agents”, says its new Action Index lets a small AI model match a frontier model on web tasks at a tiny fraction of the price. The company announced the launch on October 15, 2026. Every number below comes from Anchor’s own evaluation, so it is best read as a company claim until others test it.
What Action Index does for a computer use agent
According to Anchor, Action Index is a shared memory of web workflows that agents have already completed successfully. Instead of reasoning through every click from scratch, an agent on Anchor replays a proven path as deterministic code, and the model only steps in where a page has changed.
CEO Idan Raman put it this way: “Agents today pay to re-solve problems that are already solved.” He added that memory, not the model, is now the bottleneck for most web work, and that “the hundredth run of a workflow costs a fraction of the first.”
The benchmark numbers behind the cost claim
Anchor tested on 300 tasks from the Online-Mind2Web benchmark. A small model using Action Index completed 93.3% of tasks at an average cost of $0.016 per task. A frontier model completed 93% at $1.57 per task. Across the full 300-task run, the total cost was $4.87 with Action Index, compared with $472 for the frontier model.
The company’s headline says costs fall by 100x. By my own arithmetic, the per-task figures work out to roughly 98 times cheaper, so the headline is a fair rounding of Anchor’s own results. The release does not name either model.
The same small model, without Action Index, completed 62.7% of the tasks, which suggests the memory layer drove much of the gain in this test.
In a separate paired study on 120 tasks, Anchor says Action Index reduced 90th-percentile token use by 74%, steps by 75% and wall-clock time by 61%. Two leading open-weight models also rose from 67.7% to 88.7% and from 58.7% to 84.7%.
Availability and what is still unproven
Anchor says Action Index is available today to all Anchor accounts, including the free tier. It is offered as tools on Anchor’s MCP server and through a REST API, and the company says it works with any MCP-compatible agent or model provider.
There are good reasons for a measured reading. The results come from the company’s own evaluation, and the release does not point to an independent replication. Replaying a recorded path works best when a website behaves the same way each time, and how often pages change on real, messy websites is left open.
Security questions worth asking
Agents that log in to websites and click through accounts raise obvious security questions. A shared memory of workflows adds one more: what exactly is stored, who can see it, and how is it kept separate between customers? The release does not describe this, so I will not guess. Supporters would say deterministic code is easier to audit than a model improvising each step; skeptics would say a shared library of workflows is also a shared target. For related reading, see our piece on AI agent accountability and who gets blamed.
My view on cheaper web agents
As a writer, I feel the core idea is sensible. Reusing solved problems is how people save time, and software agents can do the same. If a small model can do the job for cents instead of dollars, that could make web automation available to far more teams. But a benchmark run by the seller is a starting point, not a verdict. I would want independent tests on messier tasks, and clear answers on data handling, before treating the savings as settled. For another look at AI moving into hard engineering work, see ChipAgents and NVIDIA testing agents on full chips.
FAQ: Anchor Action Index
What is Anchor Action Index?
It is a shared memory of web workflows that AI agents have already completed successfully. Agents replay a proven path as deterministic code, and the model only steps in where a page has changed, according to Anchor.
How much cheaper is it, according to Anchor?
On 300 Online-Mind2Web tasks, a small model with Action Index completed 93.3% at $0.016 per task, against 93% at $1.57 for a frontier model. The full run cost $4.87 versus $472.
Who can use it?
Anchor says it is available today to all Anchor accounts, including the free tier, through its MCP server and REST API, and works with any MCP-compatible agent or model provider.