> ## Documentation Index
> Fetch the complete documentation index at: https://modal-computer-use.mintlify.site/llms.txt
> Use this file to discover all available pages before exploring further.

# Latency evidence

> Map each performance claim to its status, measurement boundary, and proof.

Use this page to trace a performance claim to its narrowest source. Current results contains eligible measurements. This page also records dated, experimental, candidate, estimated, rejected, and unverified evidence.

## Current product evidence

| Claim | Status | Measurement boundary | Proof |
| - | - | - | - |
| Complete action-to-frame benchmark | Eligible | One click at `(512, 384)` through the next decoded and validated full screenshot; 100 measured samples per path | [Report](https://github.com/ashtonchew/modal-computer-use/blob/46065138902e17d2525b8a76573c4d3811064462/docs/benchmark-results-2026-08-11-provider-action-frame.md) and [artifact](https://github.com/ashtonchew/modal-computer-use/blob/46065138902e17d2525b8a76573c4d3811064462/benchmark-data/external-provider-action-frame-2026-08-11.json) |
| Same-topology Computer Step | Eligible | 100 interleaved pairs per arm through the decoded immediate frame | [Computer Step report](https://github.com/ashtonchew/modal-computer-use/blob/b60c1cb7495200e36a738c0f6e07961b1d2db93c/docs/benchmark-results-2026-08-08-computer-step.md) |
| Raw binary screenshot default | Eligible | 30 interleaved samples per arm across screenshot and pointer move | [Optimized-default report](https://github.com/ashtonchew/modal-computer-use/blob/b60c1cb7495200e36a738c0f6e07961b1d2db93c/docs/benchmark-results-2026-08-08-optimized-default.md) |
| Weighted input-work default | Eligible | Three independent mixed XTest capacity gates | [Input-capacity report](https://github.com/ashtonchew/modal-computer-use/blob/b60c1cb7495200e36a738c0f6e07961b1d2db93c/docs/benchmark-results-2026-08-08-input-capacity.md) |
| Managed standard Image lifecycle | Eligible | 30 paired lifecycles per arm | [Image lifecycle report](https://github.com/ashtonchew/modal-computer-use/blob/b60c1cb7495200e36a738c0f6e07961b1d2db93c/docs/benchmark-results-2026-08-08-image-lifecycle.md) |

## Warm-operation benchmark evidence

| Claim | Status | Measurement boundary | Proof |
| - | - | - | - |
| Warm screenshots, clicks, typing, and commands | Historical | 30 successful warm samples per path and case after target and client readiness | [Report](https://github.com/ashtonchew/modal-computer-use/blob/4425402dbc681133252dbc54d971ea4c95bc0ffc/docs/benchmark-results-2026-07-30-warm-paths.md), [Modal optimized artifact](https://github.com/ashtonchew/modal-computer-use/blob/4425402dbc681133252dbc54d971ea4c95bc0ffc/benchmark-data/modal-optimized-provider-2026-07-30.json), and [provider-default artifact](https://github.com/ashtonchew/modal-computer-use/blob/4425402dbc681133252dbc54d971ea4c95bc0ffc/benchmark-data/provider-compare-coordinate-command-2026-07-30.json) |
| Ordered action batching | Replication | 30 samples per arm for one ordered batch and four separate action requests | [Batching replication](https://github.com/ashtonchew/modal-computer-use/blob/4425402dbc681133252dbc54d971ea4c95bc0ffc/benchmark-data/modal-action-batching-ab-replication-2026-08-02.json) |
| Native XTest input direction | Replication | 30 matching Connect samples per arm | [X11 replication](https://github.com/ashtonchew/modal-computer-use/blob/4425402dbc681133252dbc54d971ea4c95bc0ffc/benchmark-data/modal-native-x11-backend-ab-replication-2026-08-02.json) |
| First changed frame | Experimental | Action completion through the first verified pixel change | [Observation artifact](https://github.com/ashtonchew/modal-computer-use/blob/4425402dbc681133252dbc54d971ea4c95bc0ffc/benchmark-data/modal-observation-2026-07-30.json) |
| Caller placement | Candidate | External caller and placed Modal Function against a requested `us-west-2` target | [Caller-placement artifact](https://github.com/ashtonchew/modal-computer-use/blob/4425402dbc681133252dbc54d971ea4c95bc0ffc/benchmark-data/modal-caller-placement-us-west-2-2026-07-31.json) |
| Command-runner tail | Candidate | 30 measured samples and one warmup per runner arm | [Subprocess-runner artifact](https://github.com/ashtonchew/modal-computer-use/blob/4425402dbc681133252dbc54d971ea4c95bc0ffc/benchmark-data/modal-subprocess-runner-ab-2026-07-30.json) |
| Fresh create to first validated screenshot | Unverified | Public create call through the first decoded and validated full screenshot | [Tracking issue](https://github.com/ashtonchew/modal-computer-use/issues/252) |
| Warm runtime cost | Estimate | Recorded resource duration combined with list prices | [Cost estimate](https://github.com/ashtonchew/modal-computer-use/blob/4425402dbc681133252dbc54d971ea4c95bc0ffc/benchmark-data/modal-optimized-provider-cost-estimate-2026-07-30.json) |
| Matched-configuration provider comparison | Unverified | Requires matched caller placement, region, resources, screenshot format, and request shape | [Tracking issue](https://github.com/ashtonchew/modal-computer-use/issues/251) |

The dated warm-operation report also supports separate-operation arithmetic. Keep that explanation with the dated evidence.

## Latency terms

| Term | Starts | Ends |
| - | - | - |
| Cold allocation | Before Sandbox creation | Modal allocates the Sandbox |
| Startup | After allocation | Daemon, desktop, auth, and required readiness complete |
| Borrow | Before lease setup | Placed client and lease are ready |
| Complete action-to-frame path | Immediately before ordered action dispatch | The next full screenshot is decoded and validated |
| First visual change | After action completion | Pixel verification confirms a changed frame or times out |
| Application readiness | Application-defined | Application-defined |

The stable Step contract ends at the immediate frame. First visual change remains experimental. Application readiness uses an application-specific signal.

## X11 shared-memory evidence

MSS remains the production screenshot source. The optional X11 shared-memory source failed its fixed promotion gate because repeated readiness measurements included capture timeouts. Keep rejected artifacts and timeout attribution with any future rerun. Promotion still requires a full passing result.

## Cost evidence

Cost estimates use recorded resource duration and list rates. A complete invoice comparison must also account for image builds, storage, network, model usage, plan fees, credits, and taxes.


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.