Anchor Launches Action Index, Cutting Computer-Use Agent Costs 100X

Anchor, the browser infrastructure company for AI agents, launched Action Index on October 15, 2026 — a shared memory of web workflows that agents have already completed successfully. Instead of reasoning through every click from scratch, a computer-use agent running on Anchor replays a proven path as deterministic code, and the model only steps in where a page has changed. The company reports the approach cuts computer-use agent token costs by up to 100X, a claim that could reshape the economics of browser-based agent automation.

The premise is straightforward: most agent web work is repetitive. Logging into dashboards, filling out forms, pulling reports from vendor portals — these are tasks an agent performs dozens of times, each time paying the full reasoning cost of a fresh attempt. According to Anchor's announcement, carried by HPCwire, Action Index records successful workflow executions into a shared index so subsequent runs become cheap replays rather than expensive re-reasoning.

Benchmark Numbers Behind the 100X Claim

Anchor evaluated Action Index on 300 tasks from the Online-Mind2Web benchmark, a widely used test suite for live-web agent performance. According to the company's reported results, a small model paired with Action Index completed 93.3% of the tasks at an average cost of $0.016 per task. A frontier model completed 93% at $1.57 per task. The full 300-task run cost $4.87 with Action Index, compared with $472 without it — roughly a 97X reduction in total spend. Notably, the same small model managed only 62.7% completion without the index, meaning the memory layer lifted both accuracy and economy simultaneously.

A second paired study on 120 tasks found that Action Index reduced 90th-percentile token use by 74%, steps by 75%, and wall-clock time by 61%. The lift extended to open models as well: two leading open-weight models rose from 67.7% to 88.7% and from 58.7% to 84.7% task completion. According to the announcement, Anchor's co-founder and CEO Idan Raman framed the shift as a move from model-limited to memory-limited web work: agents no longer pay to re-solve problems that have already been solved, and every successful run makes the next one cheaper.

Why Memory Beats Models for Computer-Use Agents

The computer-use agent category — systems that perceive screens and emit clicks, types, and scrolls to operate websites without APIs — has exploded over the past two years. Anthropic's Computer Use, OpenAI's Operator, and Google's Gemini computer-use models all tackle the same problem: turning unmodified web interfaces into operable surfaces for agents. But as analysts have noted, these agents remain brutally expensive at scale, because each task consumes large context windows full of screenshots and reasoning traces.

Action Index attacks the problem from a different angle than cheaper models or smaller context windows. It treats a completed workflow as an asset: the first successful run is expensive, the hundredth is nearly free. The replay-as-code design also addresses reliability — deterministic replays fail less often than full re-reasoning, since the agent is not gambling on fresh judgment for steps it has already proven. According to TechBullion's coverage of the launch, the model only intervenes when the page under test has actually changed, which is the standard failure point for recorded-browser-script approaches.

Availability and What It Signals for Agent Infrastructure

Action Index is available immediately to all Anchor accounts, including the free tier, exposed as tools on Anchor's MCP server and through the company's REST API. It works with any MCP-compatible agent or model provider, according to the launch materials. That compatibility matters: the agent ecosystem is fragmenting across frameworks and model vendors, so infrastructure that refuses to pick sides has a better shot at becoming shared plumbing.

The launch fits a broader pattern of browser infrastructure becoming first-class agent tooling. Anchor itself raised a $6M seed round in October 2025 led by Blumberg Capital with Gradient participating, according to Access Newswire, with Cloudflare, Coinbase, and Groq signed on as strategic partners. Founded by Idan Raman, Dor Dankner, and Guy Ben Simhon, the company positions its cloud browsers as the secure, reliable way for agents to log in, navigate, and complete tasks on websites that expose no API.

The deeper signal is economic. As agents move from demos to production, the unit economics of each task decide which use cases survive. A 100X cost reduction on repeated browser workflows — if it holds up outside Anchor's own benchmarks — would pull a wide range of back-office automations over the profitability line. The next test for Action Index is independent: whether teams running their own evaluations see the same compression, and whether shared workflow memory introduces new failure modes like stale replays on subtly redesigned pages. For now, computer-use agents just got their cheapest upgrade yet.