Ask yourself one question before you send your next prompt to Claude, ChatGPT, Perplexity, or Gemini. If a stranger handed you the output and told you it cost $500, would you believe them. Now ask the harder version. What would the same output look like if it cost $25,000.
Almost nobody asks the second question. They treat a frontier model, the same class of system four companies have collectively spent hundreds of billions of dollars training, like a vending machine. Type a question, take the first answer, move on. I have spent twenty years pricing media where the sticker price and the real cost were never the same number, and this is the same problem wearing a different outfit. The subscription price is not what you are actually paying. What you are actually paying is the gap between what the model could have produced and what your prompt asked it for. Most people are eating that gap every single day and blaming the tool when the output feels thin.
Here is what makes this worth writing about now instead of a year ago. In May, at Google I/O 2026, Google quietly rebuilt how Gemini meters usage. It dropped the old system of a fixed number of prompts per day and replaced it with what it calls a compute-based model: your usage is measured by the complexity of your prompt, the features you invoke, and the length of your conversation, refreshed on a rolling five-hour window until you hit a weekly ceiling. That is not a Google quirk. It is the same architecture Anthropic already runs for Claude, and functionally the same shape OpenAI uses for ChatGPT's rolling three-hour window on its fast model and a separate weekly allowance for its reasoning model. Four competing labs, no coordination between them, arrived at the identical design within about a year of each other. When four competitors independently converge on the same rationing mechanism, that is not a coincidence. That is what the actual cost structure of running these models looks like once the free-money era of AI subscriptions runs into the real price of compute.
Anthropic said the quiet part out loud in April. When eagle-eyed users noticed Claude Code had briefly vanished from the $20 Pro plan's feature list, Anthropic's Head of Growth, Amol Avasare, went on record with a line every subscriber to every AI tool should have to read before they sign up: "Usage has changed a lot and our current plans weren't built for this." He explained why in plainer terms afterward. Claude Max launched before Claude Code existed as a Max feature, before Cowork existed at all, and before anyone was running unattended agents for hours at a stretch. The plan was built for people typing questions into a chat box, not for the workloads people are actually running against it now. Anthropic reversed the specific test after backlash. It did not walk back the underlying admission, and a month later it struck a compute deal with SpaceX partly to buy itself room before the next round of tightening.
Read that the right way and it is not a warning to use these tools less. It is confirmation that the value on the table right now, before the next round of tightening, is larger than almost anyone is claiming.
The $500 versus $25,000 test
Here is the framework I use, and the one I'd tell any team to adopt this week. Before you accept an AI output, picture two people who could have produced it. One is a competent freelancer you found on a marketplace, charges $500 for the job, and delivers something usable but generic. The other is a senior consultant who bills $25,000 for the same category of work, the kind of person who asks four clarifying questions before starting, pressure-tests their own conclusion, and hands you something you could put in front of a board. Look at what you just got back from the model. Which one does it read like.
The uncomfortable answer, most of the time, is the $500 version. Not because the model is only capable of $500 work. Because a $500 brief produces a $500 result regardless of who or what is doing the work. If you type "write me a social media strategy" and accept the first thing that comes back, you get the freelancer. If you load the model with your actual business context, your competitive set, your unit economics, a real point of view on what you think the answer might be, and you tell it explicitly to argue with you before it agrees with you, you start getting the consultant. Same model. Same subscription. Completely different output, and the only variable that changed was how much effort you put into the ask.
This matters more than usual right now because of how the usage-metering shift changes the arithmetic. Under the old flat daily-count systems, a lazy one-line prompt and a dense, fully loaded one cost you the same unit: one prompt. Under the compute-based systems all four platforms now run, that is no longer true. A simple question burns almost nothing against your window. A genuinely hard, context-rich request burns real capacity. Which means every trivial question you fire off during the day is not neutral anymore. It is quietly spending down a budget that would have been better spent on the one problem that's actually worth $25,000 of thinking.
The strategic move nobody is talking about
Here is the part I have not seen written up anywhere, and it is the actual point of this post. If usage is metered on a rolling time window and priced by compute rather than by message count, the correct strategy is not to spread your usage evenly across the day like you're rationing water. It is to concentrate it. Treat each reset window the way you would treat an hour on a $25,000 consultant's calendar. You would not spend that hour asking them to look up something you could Google in ten seconds. You would walk in with your hardest problem, your full context already assembled, and you would use the entire hour on the one thing that actually needs that level of thinking.
Most professionals do the exact opposite. They ask ten small questions across a workday, hit a wall on the eleventh, and never get to the genuinely hard problem that was worth paying for in the first place, because they already burned the window on things a search engine could have handled for free. The fix is not more prompting. It is sequencing. Save your reset window's real capacity for the request that needs it, and route the trivial stuff somewhere it costs you nothing, a free-tier account, a lighter model, or an old-fashioned search.
Most of this can be automated, and should be. Every platform below has a way to store your standing context once so you stop paying the token cost of re-explaining yourself in every session. Use it. The businesses treating this as a workflow to design, not a chat window to type into, are the ones getting consultant-grade output at a fraction of what the actual consultant would bill. Everyone else is going to keep getting the freelancer's version and wondering why the tool feels like it's gotten worse.
Claude: batch before you send, and know what Opus actually costs you
Claude runs on a rolling five-hour session window layered under a longer weekly window, and Anthropic's own usage guidance is more direct about strategy than any other lab's. Their published best practices tell you to group multiple related questions into a single message instead of sending them one at a time, to paste entire documents for editing in one shot rather than piecemeal, and to build a Project for anything recurring, because Projects use retrieval rather than reloading your full history on every turn, which measurably extends how far your window stretches. That last point matters more than it sounds like it should. If you are re-uploading the same brand guidelines or the same client brief every session, you are paying the token cost of re-teaching Claude who you are, every single time, out of the same budget you'd rather spend on the actual work.
The other lever almost nobody manages deliberately is model selection. Opus, Claude's top-tier model, consumes somewhere in the range of five to ten times the budget per message that Sonnet does, according to Anthropic's own guidance and consistent community benchmarking. That is not a reason to avoid Opus. It is a reason to reserve it on purpose, for the handful of requests each week that are genuinely worth the consultant rate, and default to Sonnet for everything else. Pro is $20 a month for roughly 45 messages per five-hour window in typical use, which stretches to well over 200 messages a day if you space your usage instead of bursting it. Max comes in two sizes, roughly five times and twenty times that throughput, at $100 and $200 a month. Unless you are running Claude Code or Cowork against real production workloads, most people never come close to needing the top tier. They just need to stop wasting the one they already have.
ChatGPT: force Thinking mode instead of hoping for it
ChatGPT's default model, GPT-5.5 Instant, is built for speed, and OpenAI's own documentation confirms it automatically routes what it judges to be a complex question over to the slower, deeper Thinking model on your behalf. That sounds convenient. It also means the system, not you, is making the call on whether your question deserves real reasoning, and the routing logic is not something OpenAI publishes in detail. If you have a genuinely hard problem, do not leave that judgment to the router. Select Thinking manually. Plus subscribers get roughly 160 Instant messages every rolling three-hour window before falling back to a lighter mini model, and a separate weekly allowance, historically in the low thousands of manually selected Thinking messages, for the reasoning mode, which is a meaningfully larger and separately metered budget from your everyday chat allowance.
The practical rule: use Instant for drafting, formatting, and anything you'd hand to an intern. Switch to Thinking by hand for anything you'd actually pay a consultant to get right, a competitive analysis, a pricing model, a client strategy document, because that is the mode built to spend real compute checking its own work rather than pattern-matching to the fastest plausible answer. Projects and custom instructions do the same job here that Claude's Projects do: set your standing context once, and stop burning your rolling window re-explaining your business every session. At $20 a month for Plus, the gap between what you are paying and what Pro-tier reasoning could produce for you is enormous, and almost none of it gets used because most people never touch the model selector.
Perplexity: stop treating it like a search bar
Perplexity is the platform most people most badly undersell, because the interface looks like a search box and most people use it like one. The $20 Pro plan includes unlimited Pro Search and roughly 20 Deep Research queries a day, and Deep Research is not a bigger search, it is a multi-step autonomous research process that reads dozens of sources, cross-checks them, and returns a structured report with citations you can actually verify, the kind of output that would take a research analyst the better part of a day to assemble by hand. Most Pro subscribers never touch it and spend their subscription asking one-line factual questions Google would have answered for free.
Max, at $200 a month, adds Model Council, which runs your question across three frontier models in parallel and shows you where they agree and where they diverge, a genuinely useful check against any single model's blind spots when the stakes are high enough to justify a second and third opinion. It also adds roughly 10,000 monthly credits for Perplexity Labs and Perplexity Computer, enough to build and run scheduled, semi-autonomous research and reporting workflows without touching a line of code, and background Assistants that run on a schedule you set, the closest thing on this list to genuine automation you configure once and stop thinking about. If your work involves competitive intelligence, market research, or due diligence of any kind, the gap between what most people pay for Perplexity and what it is actually built to do is the widest of any tool on this list.
Gemini: the 1M token context window nobody uses correctly
Google AI Pro, at $19.99 a month, gives you something none of the other consumer tiers match at that price: a full one million token context window, roughly 1,500 pages of text or 30,000 lines of code, in a single conversation. Almost nobody uses that capacity for what it is actually good for. They ask Gemini short questions in a fresh chat every time, the same behavior pattern that wastes Claude's and ChatGPT's capacity, except here it wastes a context window most competitors do not offer at any consumer price point. The correct use of that window is to load an entire contract, a full year of financial statements, or your complete competitive research into a single conversation and interrogate the whole thing at once, which is precisely the kind of task that separates a $500 skim from a $25,000 forensic read.
As of the May 2026 shift, Google AI Pro and Ultra now run on the same rolling five-hour, compute-based window as Claude, with Ultra available at two tiers, roughly five times Pro's throughput at $99.99 a month and twenty times at $199.99, plus early access to Deep Think, Google's maximum-reasoning mode built for problems that genuinely warrant several minutes of parallel computation before answering. If your organization already runs on Google Workspace, Gemini's advantage is not the model, it is the integration, the fact that it can read a real Doc, Sheet, or Gmail thread without you exporting anything first. Most teams pay for that integration and then use it to draft emails.
The stance
Here is the part people will disagree with, and I think it needs saying plainly. Most of the "AI got worse" complaints flooding forums and LinkedIn this year are not describing a model getting dumber. They are describing a $25,000 question being asked with $500 of effort, on a rationing system now sophisticated enough to notice the difference and respond accordingly. The platforms absolutely are tightening the screws, that part is documented and I am not defending it. But the individual user's own habits are doing at least as much damage to their output quality as anything Anthropic, OpenAI, Google, or Perplexity has changed on their end. Both things are true at once, and most of the discourse only wants to talk about the first one because it is easier to blame the vendor than to admit you have been sending consultant-grade problems with intern-grade prompts for two years.
The businesses that figure this out first get a real, compounding advantage, and it is not a subtle one. If your team is treating a $20-a-month seat like a $25,000-a-month capability, because they load context once, batch their hardest thinking into deliberate sessions, and manually force the deep-reasoning mode instead of accepting whatever the router hands them by default, they are extracting value nobody has priced in yet. That gap will not stay open forever. Every lab covered here has already shown it is watching usage closely enough to change the rules the moment flat pricing stops making sense for how people are actually using these tools. Use the gap while it exists. I would rather my own team be the one extracting it than the one reading about it after it closes.