iTechGuides is reader-supported. When you buy through links on our site, we may earn an affiliate commission. As an Amazon Associate I earn from qualifying purchases. Learn more
Anchor announced Action Index on October 15, 2026, as shared execution memory for web agents. It stores successful browser workflows, retrieves them for the relevant site, and replays them as deterministic code instead of asking a model to rediscover every click. In Anchor’s benchmark, that approach delivered nearly the same success rate as a frontier model at a fraction of the cost.
What Anchor Action Index does
Action Index turns completed web-agent workflows into reusable, parameterized actions. When an agent encounters a task on a site it has seen before, Action Index can retrieve the matching workflow and execute it directly. The language model is called only when the page has changed enough that the stored action needs interpretation or adjustment.
This differs from ordinary model-driven browser automation, in which the model repeatedly inspects the page, chooses the next interaction, and pays inference cost for steps it has already solved on earlier runs. Action Index is intended to make repeated workflows cheaper as its library grows.
Replay flow
- A model-driven agent completes a workflow on a website.
- Anchor stores the successful sequence as a parameterized action rather than a one-off trace.
- Later, the agent searches the index using the site and task context.
- The matching action is replayed as deterministic code.
- If the page has changed, the model is invoked at the affected step; minor changes can use step-level fallback.
What the “100x” cost claim means
Anchor’s “100X” headline is rounded from an approximately 98-fold per-task cost difference in its technical comparison. The comparison used 300 Online-Mind2Web tasks and ran each configuration once.
| Configuration | Task success | Cost for 300 tasks | Cost per task | Approximate cost per completed task |
|---|---|---|---|---|
| gpt-5.6-luna with Action Index | 93.3% | $4.87 | $0.016 | About $0.017 |
| claude-opus-5 | 93.0% | $472 | $1.57 | About $1.69 |
The final column is simple arithmetic from Anchor’s reported run costs and success rates, not a separately measured price. The success rates were effectively tied, so the result is primarily a cost comparison rather than a claim that the cheaper setup is more accurate.
“Agents today pay to re-solve problems that are already solved. The model isn’t the bottleneck for most web work anymore. Memory is. Action Index means the hundredth run of a workflow costs a fraction of the first, and every successful run makes the next one cheaper,” said Idan Raman, Anchor’s co-founder and CEO.
Token, step and latency results
In a separate paired study of 120 Skyvern tasks, Anchor reported lower tail resource use with Action Index. The figures below are 90th-percentile reductions, meaning they describe the expensive end of the task distribution rather than an average.
Quick wins for a faster PC:
Scan for outdated or missing drivers - takes under a minuteDriver Scan →Repair Windows errors before they cause bigger problemsFix Now →| Measure | Reported reduction with Action Index |
|---|---|
| Token use | 74% lower p90 token consumption |
| Browser steps | 75% fewer p90 steps |
| Wall-clock time | 61% lower p90 elapsed time |
These results matter for production queues: reducing long-tail steps can lower timeout risk and infrastructure occupancy even when median tasks are already fast.
Effect on smaller models
Anchor also reported large success-rate gains for smaller models when they could reuse verified actions. qwen3.8-27b improved from 67.7% to 88.7%, while GLM 5.3 rose from 58.7% to 84.7%. Those results support the narrower claim that memory can compensate for some repeated browser-work reasoning; they do not establish that a small model is universally equivalent to a frontier agent on new or unfamiliar sites.
How to access Action Index
Action Index is available to all Anchor accounts, including the free tier, through Anchor’s MCP server and REST API. The documented MCP tools are:
search_action_indexto find relevant stored workflows.get_action_indexto retrieve an action definition and its parameters.execute_action_indexto run a verified action.
Anchor also provides a per-site MCP endpoint that exposes verified actions as typed tools. That lets an agent call a site-specific operation with structured inputs instead of issuing a sequence of generic browser commands.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
A practical integration pattern
- Connect the agent to the Anchor MCP server or REST API.
- At the start of a task, call
search_action_indexwith the target domain and workflow intent. - Inspect the returned action with
get_action_indexbefore execution, especially when parameters contain user or transaction data. - Run the verified workflow through
execute_action_index. - Allow the agent to handle the step-level fallback when the page differs, and record whether the action still completed successfully.
Where the savings stop
Action Index only saves the reasoning cost of a workflow that has already been solved and stored. The first run on a new site still pays the full discovery cost. A workload dominated by one-off destinations may therefore see much less benefit than a support, operations or commerce queue that repeats the same sites and tasks.
Website changes and failures
- Selector changes, redesigned layouts and A/B tests can invalidate a previously verified action.
- CAPTCHAs and proxy behavior can change what the browser sees, even when the site itself has not changed.
- Small page differences can be handled by step-level fallback, but repeated failures indicate that the action needs re-verification or retirement.
- An action that succeeds on a public benchmark may still require monitoring, permissions checks and human escalation in production.
How to judge Action Index against ordinary browser agents
A fair deployment comparison should track more than headline success. Measure success rate on the same task mix, inference cost per attempt, cost per completed task, token consumption, browser-step count and wall-clock time. Also record how many tasks are already covered by stored actions: Action Index should become more valuable as that coverage grows, while a conventional model pays its planning cost on every run.
Recovery behavior is a separate decision criterion. Test what happens after a selector change, a CAPTCHA, a slow page or an A/B-tested layout. A system that fails closed and requests re-verification is safer than one that silently continues with an outdated action.
What the launch establishes—and what it does not
Anchor’s results show a compelling benchmark trade-off: on the reported 300-task run, a smaller model paired with reusable execution memory reached essentially the same success rate as the frontier comparison at roughly one-hundredth of the cost. The evidence is still limited to public benchmark tasks, a single run per configuration and Anchor’s own implementation. Live traffic, task novelty and website volatility will determine whether the savings persist for a particular agent.
Do these 3 things before closing this tab:
1Repair Windows errors before they cause bigger problems2Fix the driver behind crashes, sound loss and screen glitches3Clear out junk files and repair common Windows errorsQuick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
