AutoBot Pricing: Free Voice Control for AI Work

AutoBot, a local harness for steering long-running AI work through native ChatGPT Voice, reached Hacker News on September 17. The launch is interesting because it combines persistent work, local memory, completion checks, and phone-to-Mac control without introducing another hosted agent subscription.

The headline pricing answer is simple: AutoBot itself is free and MIT-licensed. The full cost is not zero. It depends on the user’s ChatGPT plan, Codex allowance, Mac hardware, permissions, and any connected services. Our maintained AutoBot pricing page tracks that boundary.

AutoBot is not a model endpoint. It adds zero rows to the canonical pricing dataset and pricing API, and it does not change any existing ChatGPT subscription row.

What launched

AutoBot 0.3.0 is a public GitHub prerelease aimed at one user on a Mac. The release page provides an AutoAssist ZIP, a SHA-256 checksum, and a manifest. The project’s installation flow asks the user to verify those artifacts, open the folder as a local ChatGPT project, enable the needed native capabilities, and connect a phone through ChatGPT Remote.

The harness adds four main layers around native ChatGPT:

1. A local context and memory structure that persists across projects and conversations.
2. A durable task graph with checkpoints and a supervisor for unfinished work.
3. Explicit privacy zones and external-write checks.
4. Separate completion evidence and validation before durable work is called done.

The project does not claim to be a second model, chatbot, or cloud agent gateway. Native ChatGPT supplies Projects, Voice, Goals, Codex, Computer Use, notifications, and scheduled work. AutoBot organizes how those capabilities are used.

AutoBot pricing and the ChatGPT dependency

The project’s official setup guide says AutoBot is free but regular use will likely need a paid ChatGPT plan. The live table above renders current plan prices from our maintained subscription dataset rather than freezing them in this article.

Those are OpenAI plan prices—not AutoBot plans. Feature availability and usage still depend on the account, workspace, region, rollout, and allowance. The project also says tasks started from Voice consume the Codex usage budget. A user should therefore measure cost per completed workflow and allowance burn, not assume that a free license means unlimited agent work.

Installation makes no paid model call and enables no cloud service by itself. A supported Mac, Node 22 for the advanced runtime, storage, an always-available machine, and setup time remain real operating costs even though they are not vendor rate-card items.

What the benchmarks show

AutoBot reports two notable results:

Benchmark	Published AutoBot result	Important boundary
OSWorld 2.0	32.41% binary accuracy; 64.28% partial accuracy	Project-run best-valid-per-task aggregate with reruns
AssistantBench hidden test	50.70% accuracy; 100% answer rate on 181 tasks	Scored by the official hidden evaluator; not our run

The repository describes the 108-task OSWorld result as 18.5% higher task completion than the published OpenAI Sol Max baseline, with 35 tasks fully completed, and says it surpassed Claude Opus 5 Max. The detailed methodology matters: the final number combines 92 retained results and the highest valid result for each of 16 rerun tasks. Those tasks were rerun after the authors found that Astra had contaminated part of the intended Sol-only test.

That disclosure is useful, but the result is not a clean cost comparison. The repository does not provide one complete billed-cost ledger covering the combined run, and its desktop workflow is different from our deterministic text tasks. The AssistantBench result is externally scored, but it likewise does not establish an AutoBot subscription price or cost per accepted desktop task.

What buyers should verify

AutoBot’s architecture creates a different risk profile from a hosted SaaS agent. It runs inside one macOS user account and can use native ChatGPT capabilities that see screens, click apps, and operate connected accounts. The project explicitly says its privacy zones are procedural boundaries, not OS isolation or cryptographic identities.

Before deployment, verify:

• The exact ChatGPT plan exposes the required Voice, Codex, Remote, and Computer Use features.
• The v0.3.0 ZIP matches both the checksum and release manifest.
• The Mac permissions are limited to the apps and folders required for the first workflow.
• Sensitive accounts retain a separate fresh-verification factor the connected Mac cannot access.
• The workflow has a human-readable definition of completion and a way to inspect the actual saved or sent result.

The safest first test is a low-risk local task with two concrete files, followed by an independent readback. That exercises persistence and validation without granting access to email, finance, or production systems.

Labs decision: agent-harness and cost-ledger blocker

AutoBot cannot be inserted into the current AI Pricing Guru Labs leaderboard as if it were a model. The benchmarked product is a harness plus native ChatGPT capabilities, model settings, a Mac, permissions, and desktop applications.

A defensible comparison needs a pinned AutoBot release and ChatGPT app version; an identical plan, model, reasoning setting, Mac, and permission set; fixed OSWorld or desktop tasks; raw transcripts and tool traces; retry and operator-intervention logs; elapsed time; accepted-task grading; and actual usage, credit, or billing records. The project’s combined OSWorld methodology would also need a predeclared rerun policy for an apples-to-apples cost result.

Until that replay exists, the Labs coverage note records an explicit blocker. We report the project’s results with attribution rather than presenting them as independently reproduced.

Canonical pricing decision

The canonical model dataset remains unchanged at 230 models, and /pricing.json stays byte-identical to /api/pricing.json. AutoBot has no public model ID, token rate, credit pack, or separate paid plan, so a zero-dollar model row would be misleading.

ChatGPT subscription prices remain owned by the recurring subscription pipeline. The ChatGPT subscription pricing page is the right place to compare the plans AutoBot depends on; the dedicated AutoBot page explains what that spend buys and what it does not.

Bottom line

AutoBot is a credible open-source packaging of persistent, voice-steered ChatGPT work, not a new AI model or another agent subscription. The software license is free. The operating cost is the chosen ChatGPT plan plus hardware, setup, allowances, and connected-service risk.

The benchmark claims deserve attention, especially the official AssistantBench score, but they do not establish a universally cheaper agent. Buyers should verify entitlement and measure accepted workflows per allowance on their own Mac before treating the project as a replacement for a hosted agent platform.

Official sources: AutoBot repository, v0.3.0 prerelease, OSWorld methodology, AssistantBench evidence, and OpenAI pricing.