AutoBot Pricing 2026 — Free Harness, ChatGPT Plan Costs

What buyers actually pay for

AutoBot does not create a second usage meter. The project says tasks started from Voice also consume the Codex usage budget, and the underlying features can differ by plan, workspace, region, rollout, and permission. A free download therefore does not mean unlimited free agent work.

Buyers should budget for the ChatGPT seat first, then for a supported Mac that stays awake and online, local storage, setup time, and any external services the agent is allowed to use. The installation itself makes no paid model call and does not silently enable an API service.

Availability and release status

AutoBot 0.3.0 is a public GitHub prerelease for a single user on macOS. The release page provides an AutoAssist ZIP, a SHA-256 companion, and a manifest. The project requires users to verify the checksum before installation and warns that it is not a multi-tenant service, OS sandbox, or cryptographic privacy boundary.

Native Voice, remote phone control, Computer Use, Codex, connectors, and notifications are separate ChatGPT capabilities. AutoBot cannot unlock a feature, bypass a plan limit, or broaden approval authority. That distinction is important when comparing it with a standalone hosted agent product.

How to read the benchmark claims

The repository reports 32.41% binary accuracy and 64.28% partial accuracy on OSWorld 2.0 with GPT-5.6 Sol Max, plus 50.70% accuracy and a 100% answer rate across 181 hidden AssistantBench tasks. These are project-reported or externally scored results, not AI Pricing Guru Labs runs.

The OSWorld figure is a best-valid-per-task aggregate that combines 92 retained results and reruns 16 tasks after the authors found an Astra contamination issue. That is useful methodology disclosure, but it is not a single clean cost-comparison run and the release does not publish a complete billed-cost ledger that can be normalized against our 49-task text suite.

Why AutoBot is not in the pricing API

The canonical pricing dataset and public API track billable model endpoints. AutoBot is a software harness with no model ID, per-token rate, or separate subscription SKU. Adding a zero-dollar model row would imply a callable free model that does not exist.

Existing ChatGPT plans remain in the maintained subscription dataset. No model or subscription row changed because of this Show HN launch.

Labs inclusion decision

AutoBot is not a model route, so it cannot be dropped into the current Cost-per-Task leaderboard. A valid agent-harness comparison needs a pinned AutoBot release, identical Mac and ChatGPT plan, fixed base-model and reasoning settings, matched OSWorld or desktop tasks, raw transcripts, tool traces, retries, operator interventions, elapsed time, accepted-task grading, and actual usage or credit records.

Until that replay exists, Labs records an explicit agent-harness and cost-ledger blocker rather than copying the project’s headline scores into our leaderboard.

Buyer checklist

• Confirm that the required ChatGPT features are enabled for the exact plan, workspace, region, and account.
• Verify the release ZIP against its published checksum and manifest before installation.
• Start with a narrow project and least-privilege app permissions.
• Measure completed workflows per allowance, not just the software’s $0 license price.
• Keep high-risk accounts and recovery factors outside the connected Mac profile.

Read the launch analysis for the Show HN context and benchmark caveats.