ChatGPT dots pricing clarified: the $100/$200/$500 tiers and Rakazo's real costs

A post on dev.to published October 2, 2026 takes aim at a pricing claim circulating in coverage of AI teammates: that OpenAI's ChatGPT dots costs $100 a month while the open-source alternative Rakazo is free. According to the author, that comparison conflates two unrelated numbers — the price of a ChatGPT subscription tier and the cost of running your own infrastructure.

The $100 figure is a Pro tier, not a per-agent fee

The post states that as of October 2, 2026, $100 is the entry-level monthly price for ChatGPT Pro, which also comes in $200 and $500 tiers. Citing OpenAI's documentation, the author says a first dot is included at no extra charge in eligible Pro and Business Premium plans. The $100 figure people quote is therefore the price of the plan the agent rides on, not a separate charge for the agent itself.

What self-hosting actually costs

Rakazo's Apache-2.0 license removes software licensing fees and nothing else, the post argues. Hosting, model inference tokens, external sandbox services and maintenance time all still cost money, and the author frames the monthly bill as server hosting plus model API invocations plus sandbox services plus the time spent keeping the stack running. There is also a data-boundary caveat: even with the bot on hardware you control, calls to third-party commercial models still transit external networks, so self-hosting the runtime does not self-host the inference.

Managed sandbox versus your own runtime

On the managed side, the post describes dots as an agent operating inside an OpenAI-provided cloud computer and browser, using GPT-6 Astra within the ChatGPT ecosystem, with what the author claims is a catalog of more than 4,000 plugins. Costs are bundled into the Pro and Business Premium tiers, but an always-on agent does not mean unlimited deep inference — the post says dots operates under usage limits set at the account level.

Rakazo takes the opposite trade. You run it wherever you like — local Docker, a VPS, or sandbox providers such as E2B, Daytona or Box — and point it at any model provider through Pi or OpenAI-compatible APIs. Integrations come via Composio, Pipedream Connect, MCP servers or OpenAPI specs rather than a single vendor catalog.

Approval gates on both paths

The post reports that OpenAI enforces approval gates for financial transactions and password management, and claims Rakazo's verification test suite confirms that destructive computer actions wait for human approval before executing. In the author's framing, both platforms treat irreversible actions as human decisions rather than agent defaults.

Spec first, autonomy later

Much of the post is a workflow for defining an agent before deploying one: draft a structured task card in one prompt, then audit it in a second for edge cases — what happens on day one when no baseline exists, and how the agent distinguishes a transient network timeout from genuinely having no updates to report.

The resulting task contract specifies a trigger time with an explicit timezone, scope limited to canonical documentation and release feeds, comparison against the previous successful run, bounded output of at most three validated items with timestamps and direct URLs, and circuit breakers that halt the agent when it hits authentication walls, CAPTCHAs, contradictory data or write actions.

The rollout advice is deliberately slow: run the routine for three consecutive days on public data, confirm URLs resolve and false positives are gone, and hold off on linking production Slack channels or primary email accounts until human review is in place for anything the bot sends outward.

Why it matters

Comparisons between managed and self-hosted AI agents usually stop at sticker price: a subscription versus free software. This post lays out the comparison engineers actually face — predictable subscription cost with usage caps on one side, unbundled hosting plus token spend plus operations time on the other, with different data boundaries and approval models on each path. As teammate-style agents become line items in engineering budgets, the distinction between what a Pro tier includes and what a daemon costs to run is the part worth getting right. The claims here come from a single community post, so the tier pricing and dot inclusions are worth checking against OpenAI's current documentation before committing to either option.