Audit AI feature pricing: free core, meter heavy work, upgrade at limit

Your AI feature's free tier is the first thing users judge. One Cursor customer received a $7,225 invoice after one developer used 500 requests in a single day. Keep four moves in the audit: define the free everyday AI surface, meter only compute-heavy generation, set a wide tier spread, and fire the upgrade prompt at the limit.

The free core is the trust surface

Designers often treat the free tier as a trial. Users treat it as the product. If the default path is blocked by a paywall, the feature feels like a vending machine rather than a tool.

The core should cover the routine task: asking a question, drafting a short reply, summarizing a page, or checking a basic result. Meta's core experience across its apps and Meta AI remains free, while subscriptions add specialized capabilities and expanded AI usage. The free surface should be the daily habit that makes the paid headroom worth buying.

For a social or creator product, the free core can include the basic assistant, the standard filters, and the ordinary search. The paid layer should add the things that cost more to run or require more curation.

Meter heavy work and spread the tiers

Meta introduced Meta One on Tuesday, Sept. 15, to monetize its AI efforts. It launched with more than 50 features across Meta's social media and AI platforms, with planned expansion to Edits, AI glasses, and other offerings. A broad bundle makes the metering question harder: which features are everyday, and which are headroom?

Every AI feature has light work and heavy work. Light work is quick, bounded, and cheap enough to absorb. Heavy work is long, recursive, multimodal, or batched, and it can eat a margin in a session.

The meter should follow compute. A short answer can be free. A long document rewrite, a video edit, a large image batch, or a multi-step agent run can be paid. If a user can predict the cost, the meter feels fair. If the cost appears after the fact, it feels like a trap.

Cursor's June 2025 move from a flat 500-request monthly plan to credit pools provoked a developer backlash, and CEO Michael Truell issued a public apology on July 4, 2025.

In February 2026, Anthropic's Claude and OpenAI's ChatGPT both had free tiers and paid tiers, with Claude carrying $17, $100, and $200 monthly plans and ChatGPT carrying $20, $30, and $200 monthly plans. The shape matters: a low entry, a mid tier for regular heavy use, and a top tier for power users.

Meta One's individual bundles include a $7.99/month Core plan and a $19.99/month Premium plan, and each plan includes Facebook Plus, Instagram Plus, and WhatsApp Plus features. Its creator and business tiers start at $14.99/month for Essential, $49.99/month for Advanced, $149/month for Expert, and $499/month for Max. The wide spread gives users a reason to start low and a path to move up when their usage grows.

The upgrade prompt belongs at the limit

Most upgrade prompts fire too early. They interrupt the user inside the free core, which makes the product feel salesy. The better moment is the limit: the user has just tried to do something real, and the system has a clear reason to stop.

The prompt should be precise. It should name the limit, show what the paid tier unlocks, and let the user continue without shame. A message like You have reached your limit for long generations is better than Upgrade now to get the best experience. The first is a receipt. The second is a banner.

A daily cap should show its reset time, a monthly cap its remaining amount, and a compute cap the work that consumes it. Vague limits create distrust. Clear limits create a reason to buy.

The audit belongs before launch

Use this checklist before you ship an AI feature or change its pricing. Each item maps to a trust failure: a blocked core, a broad meter, a steep jump, a vague limit, an early prompt, or a hard exit.

• Free core: Can a new user complete the main task without paying? A no answer means the core is too small.
• Metered work: Does the paid meter attach only to long, batched, multimodal, or agent-style generation? Taxing short answers makes the meter too broad.
• Tier spread: Is there a low entry, a mid tier for regular heavy use, and a top tier for power users? A steep jump stops users from moving up.
• Limit clarity: Does the UI show the limit before it hits, and does the copy explain what consumes it? Guessing makes the limit feel arbitrary.
• Upgrade prompt: Does the prompt appear at the limit, name the limit, and offer a clear next step? An earlier prompt feels like a sales interruption.
• Trust check: Can a user cancel, downgrade, or return to the free core without friction? A hard exit makes the paid headroom feel like a trap.

Re-run it after the next wave of usage data.