For three years, paying for AI was simple: $20 a month, use it as much as the limits allow, done. In 2026 that model started to crack. The most capable models are moving behind usage-based billing, and the plans that still look flat are rationing their best features more tightly.
What Changed This Year
Top models went metered. Since July 20, Claude Pro and Team Standard subscribers reach Anthropic's flagship Claude Fable 5 through usage credits, billed at $10 per million input tokens and $50 per million output tokens. Max and Team Premium plans still include it, but only up to 50% of their usage limits.
Limits became compute-based. Google moved Gemini away from a fixed number of prompts per day toward allowances based on how much compute you use. A long, complex request now spends more of your allowance than a short one.
Plans split into tiers. The competition moved to two price points: a budget tier under $10 (ChatGPT Go at $8, Google AI Plus at $4.99) and a power-user tier around $100 a month. The $20 plan is still there, but it no longer covers everything.
The common thread: what you pay increasingly depends on how many tokens you burn, not just which plan you picked.
Why Retries Are Now a Line Item
On a flat plan, a bad prompt costs you time. On a metered plan, it costs you money too.
Think about a typical exchange. You ask a vague question, get an answer that misses the point, clarify, get something closer, then ask for a rewrite in the right format. That's three full round trips — and on metered billing, you pay for every token of all three. The two answers you threw away cost exactly as much as the one you kept.
Output tokens matter most. At $50 per million output tokens, a model writing long answers you don't need is the most expensive thing that can happen in a session. Asking for exactly the length and format you want is one of the simplest ways to cut cost.
Five Habits That Keep Metered AI Affordable
1. Say what "done" looks like. State the format, length and audience in the first message. "A five-bullet summary for a non-technical manager" gets you there in one try; "summarize this" often takes three.
2. Front-load the context. Paste the relevant details up front instead of drip-feeding them across follow-ups. Each follow-up resends the whole conversation, so late context costs more than early context.
3. Cap the output. If you need a paragraph, say so. Long answers to short questions are pure waste on a meter.
4. Start fresh for new tasks. A long chat carries its full history into every new message. When the topic changes, a new conversation is cheaper.
5. Use the flagship only when it earns it. Simple jobs — reformatting, tagging, short summaries — rarely need the most expensive model. Save it for the hard reasoning.
Predictable Costs Start With the Prompt
You can't control what AI providers charge. You can control how many attempts each task takes and how many tokens each answer uses. Those two numbers decide your bill far more than the plan you pick.
IacuWise rewrites your prompt before it reaches the model — adding structure, context and output constraints — so you get a usable answer on the first try. Paste a prompt, pick your model, and see the optimized version alongside the tokens it saves.