Why a subscription is not enough for a constantly running AI agent

The subscription ran out before the agent finished the job

The main issue is clear: a monthly price says almost nothing about how much agent work is actually available. In the cited user discussion, a $20 ChatGPT subscription reportedly hit a five-hour limit after 30 minutes, while the same task through Antigravity took five minutes and barely affected the limits.

This is the only primary source for the case described, so these figures should be treated as user experience rather than an official benchmark. The thread provides no prompts, repository size, tool-call counts or model settings. Without that information, the speed comparison is illustrative but not reproducible.

Another participant says a $100 subscription was enough for full-time work with Sol 5.6, whereas Astra burns through a weekly limit in a day. Several other estimates are harsher: active generation allegedly requires two $200 subscriptions, and without limit resets the discussion turns to three such subscriptions.

Corporate Anthropic access is mentioned separately: one user estimates average Opus spending at $2,000–$3,000, excluding Fable. This is neither a published price nor a verified bill, but a private estimate from the same thread. As of September 11, 2026, these amounts cannot be treated as a current price list because neither the discussion date nor official limit tables are provided.

Technically, such variation is easy to explain. An agent spends its budget on more than a single answer: it repeatedly reads context, plans, calls tools, checks results and makes fixes. High-effort settings and fast mode can sharply change throughput even within the same subscription.

Measure completed tasks, not subscriptions

The practical conclusion is simple: for an autonomous coding agent, cost per completed task and time until throttling matter more than the plan price. A cheap subscription loses if the agent loops for 30 minutes and then stops until its limit resets.

I would first measure the number of agent steps, the amount of context repeatedly sent back, the share of useful changes and waiting time. Then I would test long sessions separately: that is where expensive rereads, failed branches and repeated tool calls accumulate.

This thread does not prove that one model is unconditionally more economical than another. But it clearly shows the gap between an interactive assistant and a continuously operating agent: without a transparent compute budget, the word “subscription” hides the system's key parameter—how much useful autonomous work it can actually sustain.

We previously examined how the cost of Claude Opus 4.6 depends on context, reasoning mode and the selected configuration. This helps compare plan limits with the real costs of running AI agents.