Anthropic, OpenAI, and Perplexity Quietly Shift to Costly Auto-Reload as AI Goes Agentic

With the advance of agentic AI tools, Anthropic, OpenAI, and Perplexity have established a new billing model: features that automatically purchase additional credits once the credit budget included in a subscription has been used up. Anthropic calls this “Extra Usage” with “Auto-Reload,” OpenAI refers to it as “Automatic Reload” or “Auto Recharge,” and Perplexity calls it “Auto-Refill.”

The mechanics are similar in all three cases: a stored payment method is charged as soon as a defined minimum balance is reached — without requiring confirmation for each top-up. What is positioned as a convenience feature to prevent workflow interruptions quickly becomes a cost trap with agentic workloads. With more intensive use of agentic capabilities in particular, which consume far more tokens than ordinary search queries, the credits included in a subscription are quickly exhausted — and to continue working, the AI then demands new credits. Or simply picks up again tomorrow.

Perplexity: Auto-Refill with a cap of up to $2,000

Perplexity Max costs $200 per month and includes 10,000 credits per month as well as a one-time launch bonus of 20,000 credits that expires after 30 days. The Auto-Refill feature is disabled by default. When activated, Perplexity automatically purchases additional credits as soon as the balance falls below 500 credits (approximately five dollars). The default monthly cap is $200. In addition, there is a separate account spending cap that defaults to $200 per billing period and can be raised to up to $2,000. Unused monthly credits expire at the end of the billing period.

How quickly agentic tasks can blow through the standard budget is illustrated by a run documented in the community: a Reddit user asked Perplexity Computer to check a 280,000-line Python codebase for bugs. The run took around 40 minutes and initially consumed 15,000 credits, eventually climbing to 21,000 credits, with a further 2,000 credits burned in an attempt to push the result to GitHub. That amounts to more than double the standard monthly allocation — for a single task.

Anthropic: Extra Usage with a daily limit of $2,000

At Anthropic, the “Extra Usage” feature allows users on Pro, Max, Team, and Enterprise plans to continue working at API rates once the limits included in their subscription have been reached. According to the Help Center, the feature includes a monthly spending cap, Auto-Reload settings that automatically purchase credits when a threshold is reached, and usage alerts. The daily redemption limit is $2,000. Extra Usage applies uniformly to Claude, Cowork, and Claude Code.

For Team and Enterprise plans, the logic is structured at the organizational level: owners can enable Extra Usage in the Organization Settings, activate Auto-Reload with a defined minimum balance, and set an organization-wide monthly spending limit. For Claude Code, Anthropic notes that Auto-Reload is managed not through the terminal tool but in the Claude Console account, and that all transitions to API credit usage require explicit user consent.

However, cases have been documented in the community where the billing logic did not function as described. A bug report filed on GitHub by a Pro user describes how their Extra Usage budget was drained to zero euros, with €5.91 charged against a €5 monthly limit — even though the plan limits still showed available capacity and Auto-Reload was disabled.

OpenAI: Auto-Recharge active by default in API setup

At OpenAI, the Auto-Reload logic exists across multiple product tiers. Plus and Pro users of Codex can enable Auto Top-up in Settings under Usage. If the credit balance falls below the chosen minimum, only the amount needed to bring it back up to the target balance is automatically purchased. Auto Top-up applies to the shared credits balance across Codex and Sora.

For ChatGPT Business, the feature is implemented at the workspace level. Automatic Reload is designed to prevent interruptions when a workspace consumes credits for Codex activity. Users set a minimum balance and a target balance. An optional monthly recharge limit can be used to cap Auto-Reload purchases — if the field is left empty, unlimited automatic top-ups per month are possible. Codex seats in ChatGPT Business have no fixed monthly seat fee but require workspace credits for every activity.

With API integration, the default behavior differs from consumer plans: Auto Recharge is enabled by default during the initial setup. Anyone who does not want automatic top-ups must actively deactivate the option before confirming the initial credit purchase. OpenAI also documents a limitation in its own billing accuracy: due to the complexity of billing systems, there may be delays before access is cut off after all credits have been consumed. This excess consumption appears as a negative credit balance and is deducted from the next credit purchase.

Why agentic AI is changing the economics

These features are no coincidence — they are a direct response to the architectural shift in the tools. As long as AI functioned as a chatbot — question, answer, a few thousand tokens — consumption was predictable. Agentic systems work autonomously over longer periods, break tasks down into sub-tasks, and execute them in parallel. Claude Code sessions can exhaust the five-hour window of a Max subscription in under 90 minutes. Multi-agent setups with three or four coding agents running in parallel consume not a multiple of a single chat, but an order of magnitude more.

GitHub has already drawn conclusions from this dynamic and is switching Copilot to usage-based billing with AI Credits on June 1, 2026 — with the explicit justification that agentic use “is becoming the default” and that the flat-fee model can no longer withstand the compute demands. Sam Altman has publicly stated that OpenAI must evolve into an “AI inference company” — an acknowledgment that agentic use requires a fundamentally different economic model.

Structurally, this mechanism differs from classic pay-per-use models because the system itself decides on the depth — and therefore the cost — of execution, rather than the user alone deciding through the frequency of their requests. This is precisely where the cost risk of Auto-Reload lies: a single misdirected agent loop, an overly broadly defined task, or a multi-agent setup that runs into a dead end can exhaust the monthly cap within minutes.