Most of what your team runs through AI is routine work that a cheaper model handles just as well, and this week's pricing pages make it easy to prove it.
Here's a bet: somewhere on your team this week, someone pointed the priciest model on your AI plan at something dumb. Summarizing a call transcript. Cleaning up a Slack message. Tagging a CRM record. That's the AI equivalent of hiring your most senior specialist to alphabetize a filing cabinet. By the end of this you'll know the real numbers behind why that's wasteful, and how to stop doing it in about ten minutes, using a free tool that runs in a browser tab.
The idea, in plain language
It's called model tiering, or cost-based routing, and it's simpler than it sounds. Not every AI model costs the same to run, and the expensive ones aren't expensive because someone padded the price. They're bigger, slower, and built to reason through the hardest problems you can throw at them. A frontier model is the senior specialist. A smaller, faster model is the capable generalist who handles the other ninety percent of your workload for a fraction of the cost. The skill isn't picking one model forever. It's matching the model to the task, every time you run one.
None of this matters if your team sends one AI prompt a day. It matters once a task repeats. Every meeting gets transcribed, every support ticket gets triaged, every outbound email gets a first draft, every CRM record gets tagged, and the price difference between models is charged per token. That means it multiplies with exactly the kind of volume a growing team generates.
What the pricing pages actually say
Look at what Anthropic charges by the token, laid out on its models overview page. At the bottom of the current lineup, Claude Haiku 4.5 lists at $1 per million input tokens and $5 per million output tokens. Claude Sonnet 5 lists at $3 and $15 (discounted to $2 and $10 through the end of August 2026). Move up to Claude Opus 5 and you're paying $5 and $25. Move up again to Claude Fable 5, the most capable model Anthropic ships, and it's $10 and $50, ten times Haiku's output rate for the same million tokens. None of that is a knock on the expensive models. It's the point: they cost more because they're built to do more, and if your task doesn't need that headroom, you're paying for it anyway.
The same math shows up outside Anthropic. Moonshot AI's Kimi K3, a 2.8 trillion parameter open-weight model released July 16, 2026, lists on OpenRouter at $3 per million input tokens and $15 per million output tokens, the identical rate as Claude Sonnet 5. That's worth sitting with. Two frontier-class labs landed on the same sticker price for their mid-tier offering in the same month. The market has started pricing "capable enough for most things" as its own category, distinct from "the absolute best available." That category is exactly where your routine work belongs, the meeting recap, the first-pass reply, the batch of product descriptions that all need the same tone.
Where you actually go to do this
Here's where it gets usable without an engineer. OpenRouter's model list is a browser page that puts essentially every commercial and open model side by side, with live pricing, context length, and benchmark scores you can sort and filter. Filter by price, sort by the benchmark that resembles your task, and you'll see the frontier models sitting near the top of the price column and a long list of cheaper options that still land respectably on the same benchmarks. Open the page for a specific model, like Kimi K3's listing, and there's a Playground link where you paste in a real task and read the output before you commit a workflow to it. No terminal, no API key required just to look, nothing to install.
Compare that to what engineers have been building for the same problem. OmniRoute, a self-hosted, MIT-licensed AI gateway on GitHub, has crossed 34,000 stars doing this same routing at the infrastructure level: one endpoint that talks to 290-plus providers and 500-plus models, with automatic fallback when one runs out of quota. It's a genuinely useful project, and it's also not something a marketing lead installs on a Friday afternoon. It wants Docker, environment variables, and a tolerance for reading GitHub issues. OpenRouter's model list gets you most of the same insight, the actual price and performance comparison, without any of that setup, and it runs in the browser you already have open.
Where this goes wrong
Here's the pitfall. People see a benchmark table, notice the cheaper model scores close to the expensive one, and flip an entire workflow over in one move. Don't do that. Kimi K3's own published benchmarks come with a wall of footnotes: different scores depending on which coding harness ran the test, which reasoning effort setting was used, whether the task hit a content-safety fallback. Benchmark numbers don't transfer cleanly to your specific task. A model that scores well on a public reasoning test can still miss the detail that matters in your actual CRM export. Context window matters too. Claude Haiku 4.5 tops out at a 200,000-token context window, while Sonnet 5, Opus 5, and Kimi K3 all handle a million tokens, so the cheapest option isn't always the right fit if your task is stuffing an entire contract or a quarter's worth of transcripts into one prompt. And reasoning models like Kimi K3 run with "thinking" turned on by default, which means a cheap-looking per-token price can still add up to a large bill if the model reasons through a lot of extra tokens to get there. Test the swap on a real task with real output before you trust it with anything that touches a customer.
Try this tomorrow morning
Open OpenRouter's model list this week, pick one recurring task your team already runs through your most expensive AI plan, meeting notes, email drafts, ticket tagging, and run the identical prompt against two or three cheaper models in the playground. Judge the output against what your frontier model produced, not against a benchmark score. If a $3-per-million-token model does the job as well as a $50-per-million-token one, move that task there and save the expensive model for work that actually needs it. Write down which task you moved and what it used to cost you in time or a contractor's invoice, because that's the number that will convince the rest of your team this is worth doing. If it doesn't hold up, you've lost ten minutes and learned something real about your own workflow.
The frontier model isn't your assistant for everything. It's the one you call when the task is actually hard, and the pricing pages now make it embarrassingly easy to tell the difference between hard and just tedious.