Anchor announced Action Index on October 15, 2026, as shared execution memory for web agents. It records successful workflows as parameterized actions, retrieves them by website domain, and replays them as deterministic code instead of asking a model to rediscover every click.
The “100×” headline is a rounded description of Anchor’s approximately 98× per-task cost comparison. It refers to the cost of completing benchmark tasks, not a universal 100× reduction in token consumption; a separate study measured a 74% reduction in 90th-percentile token use.
As an Amazon Associate I earn from qualifying purchases.
What Anchor Action Index does
Traditional browser agents repeatedly reason through the same login, checkout, search, or form-filling sequence. Action Index is designed to make that repetition cheap. After an agent completes a workflow, Anchor stores the successful sequence as a parameterized action. When a later request targets the same site, the system searches its index by domain, retrieves a matching action, and executes deterministic code for the known parts of the workflow.
The model is called again only when the page differs enough that the stored action needs interpretation or adjustment. That division lets a smaller model handle occasional changes while routine steps avoid a fresh reasoning pass.
#1 Best Overall
Anchor co-founder and CEO Idan Raman described the rationale this way: “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.”
What the 100× cost claim actually measures
Anchor compared two configurations on the same 300 Online-Mind2Web tasks. The Action Index configuration used gpt-5.6-luna; the comparison used claude-opus-5 without the same indexed execution path. Success rates were effectively tied, while the reported per-task costs differed sharply.
| Configuration | Task success | Reported cost per task | Approximate cost per successful task* | Spend for 300 tasks |
|---|---|---|---|---|
| gpt-5.6-luna plus Action Index | 93.3% | $0.016 | About $0.017 | $4.87 |
| claude-opus-5 | 93.0% | $1.57 | About $1.69 | $472 |
*The cost-per-successful-task figures are arithmetic using Anchor’s reported cost and success rate, not separately reported measurements.
Free tools Windows power users keep installed
One-click scans. No signup required.
Anchor characterizes the difference as a 98× reduction in cost per task. The company’s 100× wording is therefore a rounded marketing frame, not an exact multiplier. The evaluation was a single run for each configuration, so it establishes the result for that benchmark run rather than guaranteeing the same ratio in production.
Token, step and latency savings in a second study
A separate paired study of 120 Skyvern tasks measured how indexed execution changed the expensive tail of agent runs. Anchor reported the following reductions at the 90th percentile (p90), meaning the point below which 90% of observed runs fell:
| Metric | Change with Action Index |
|---|---|
| Token use | 74% lower p90 usage |
| Number of steps | 75% fewer p90 steps |
| Wall-clock time | 61% lower p90 duration |
These figures explain why the headline should not be read as “100× fewer tokens.” The strongest token result in the supplied evaluation is a 74% p90 reduction, while the approximately 98× figure compares total task cost between two configurations.
Can a smaller model approach a frontier agent?
Anchor also tested whether indexed actions can raise the success rate of less capable models. On the reported tasks, adding Action Index improved both models substantially:
Crashes, No Sound, or Screen Glitches?
Random freezes, missing sound and display glitches usually trace back to one bad driver. Find and replace yours safely.Free scan · under a minutePC Slower Than It Used to Be?
A free scan shows the junk files, broken settings and background clutter dragging Windows down - then fixes them in one click.Free scan · Windows 10 & 11| Model | Success without Action Index | Success with Action Index | Improvement |
|---|---|---|---|
| qwen3.8-27b | 67.7% | 88.7% | 21.0 percentage points |
| GLM 5.3 | 58.7% | 84.7% | 26.0 percentage points |
That result supports a narrower conclusion than “a small model becomes a frontier model.” Indexed workflows can supply the model with a verified procedure for previously solved sites, reducing how much open-ended browsing it must do. The advantage depends on finding a relevant action and on that action still matching the live website.
How to use Action Index through MCP or the REST API
Action Index is available to all Anchor accounts, including the free tier. Anchor exposes it through its MCP server and REST API. The documented MCP workflow is:
Rank #4
- Connect your agent. Configure the agent to use Anchor’s MCP server, or make requests through the REST API.
- Search for a known workflow. Call
search_action_indexwith the target website domain and the operation the agent needs. - Retrieve the action. Use
get_action_indexto obtain the matching parameterized workflow and its required inputs. - Execute the verified procedure. Call
execute_action_indexto replay the action. A per-site MCP endpoint can expose verified actions as typed tools, allowing the model to invoke a site-specific operation directly. - Handle exceptions. If the page has changed, let the agent use step-level fallback to interpret the changed portion rather than abandoning the entire workflow.
In practice, the first successful run on a new site still pays the full reasoning cost. The economic benefit appears on later runs after a usable action has been stored.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.Where the savings stop
Action Index is most useful for repeated workflows with stable interfaces. It is not a guarantee that any arbitrary browser task will become inexpensive.
- New sites and unsolved tasks: there is no stored procedure to replay, so the initial run must reason through the workflow normally.
- Website changes: selector changes, redesigned pages, A/B tests and altered navigation can invalidate an action. Step-level fallback can accommodate small changes.
- CAPTCHAs and access controls: these can interrupt deterministic execution and require a separate handling path.
- Proxies and runtime conditions: network behavior or browser-environment differences can change whether a previously verified action succeeds.
- Repeated failures: an action that no longer works should be re-verified or retired rather than replayed indefinitely.
Anchor’s benchmark tasks were public datasets, each configuration was run once, and the results may not predict production traffic with changing websites, different task mixes or higher concurrency.
Best Value
What the launch means for browser-agent design
The architectural shift is from treating every browser task as a fresh model session to treating successful procedures as reusable software artifacts. A model remains responsible for discovering a workflow, supplying parameters and recovering from meaningful changes. The index handles the deterministic middle of a workflow at much lower recurring cost.
Teams evaluating the approach should therefore measure more than raw success rate. The useful comparison is success on previously solved workflows, cost per task and per completed task, token use, step count, wall-clock time, and recovery behavior after a website changes. Those measurements reveal whether an agent is benefiting from accumulated execution memory or merely paying to browse the same pages repeatedly.
Quick 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.
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 errors
