AI Model Upgrade Advisor

What Is the AI Model Upgrade Advisor?

The AI Model Upgrade Advisor is a free tool that finds cheaper AI models than the one you are using today. You tell it your current provider and model, how many input and output tokens a typical request uses, and how many requests you send per month. It then scans the full models.dev pricing catalog and returns the alternatives that cost less for exactly that workload, ranked by monthly cost.

The word upgrade here means a smarter purchase, not necessarily a bigger model. Sometimes the cheapest move is down to a smaller model, sometimes sideways to another provider with the same capabilities, and sometimes the answer is that your current model is already the best value, which the advisor will tell you plainly.

What makes it different from a spreadsheet is the compatibility logic. A model that is cheaper but cannot call your tools, or cannot fit your prompts in its context window, would break your product. The advisor filters those out for you by default, so every suggestion is a switch you could actually make.

Why Use It — and Who It Is For

AI bills grow quietly. A model that looked affordable at prototype volume can dominate your costs at production volume, and providers change pricing and deprecate models often enough that last year choice deserves a second look.

• Engineering leads reviewing a growing monthly AI invoice and looking for savings that do not break the product.
• Founders and product managers trying to extend a runway without cutting features.
• Developers inheriting a legacy integration still pointed at a discontinued or overpriced model.
• Consultants who need a defensible, numbers backed recommendation to hand to a client.
• Indie hackers and students running personal projects on a fixed budget.

The typical question is: I pay for model X, what else works for my workload and how much do I save? This tool answers it with current catalog prices instead of blog posts from last year.

How This Tool Works

The page loads the provider list from our backend, which reads the models.dev catalog that our server downloads, caches for 24 hours and stores a local copy of. Choosing a provider loads its models, and picking a model shows its badges: provider, status, reasoning, tool calls, open weights and context size. If the model is discontinued, a warning banner appears above the results area.

When you press Get Recommendations, your current model, usage numbers and both switches are sent to the recommend endpoint. The server then runs this pipeline:

1. Baseline. Your current model is located in the catalog and priced with your usage. If it has no published rates, there is no baseline to compare against, and rows come back without a saving figure.
2. Validation. At least one of input or output tokens must be above zero, otherwise you get the message "Enter at least one token count greater than zero."
3. Candidate filter. Every other model in the catalog is checked. Discontinued models are never suggested, and models with no published input and output prices are skipped.
4. Compatibility filter. With strict compatibility on (the default), a candidate must have the same reasoning support as your current model, must support tool calls if your current model does, and must offer a context window at least as large as the one you rely on. The required window is based on a floor of 2000 tokens, raised to your input size, and capped by your current context size. With strict mode off, only that input size floor is enforced.
5. Pricing and savings. Each surviving candidate is priced with your exact usage. Savings are calculated as a percentage of your current monthly cost, rounded to one decimal place. Candidates that cost more than your current model are discarded, unless you turned on the deprecated switch for a discontinued model.
6. Ranking. Results are sorted by monthly cost, cheapest first, and the top three are returned together with the total number of alternatives found.

Everything is priced in USD per 1 million tokens from the live catalog, and every response reports when that data was last updated.

Key Features

• Provider and model pickers with a filter box, so you can find your model by typing part of its name.
• Model badges showing status, reasoning support, tool call support, open weights and context size before you run anything.
• Usage inputs for input tokens, output tokens and requests per month, which is all the pricing math needs.
• Strict compatibility switch (on by default) that keeps reasoning, tool calling and context window requirements intact.
• Deprecated model switch that allows replacements which cost the same or more, for moving off a discontinued model.
• Deprecated banner that appears automatically when your current model has been discontinued.
• Hero comparison showing your current model, the suggested switch, both monthly costs, the saving percentage and the dollar amount saved per month.
• Alternatives table with input and output rates, per request cost, monthly cost, savings badge and a switch button for each row.
• Save and move badges - a green badge with the percentage saved, or an amber recommended move badge when the point is continuity rather than savings.
• Monthly cost chart comparing your current spend against every suggestion at the same volume.
• Use this button that swaps a suggestion in as your current model and reruns the analysis, so you can chain several upgrades in one session.
• Honest empty state that tells you no cheaper alternative exists with the current settings and suggests loosening compatibility or changing volume.

How to Use — Step by Step

1. Open the Current Model panel. The first provider in the catalog is already selected for you.
2. Change the Provider if needed. The model list reloads for that provider.
3. Type part of a model name in the filter box to narrow the list, then select your Model. Read the badges that appear underneath.
4. If the model shows a deprecated badge, note the warning banner that appears in the results column.
5. Enter Input tokens and Output tokens per request. Defaults are 1000 and 500, which suit a small chat style request.
6. Enter Requests per month. The default is 10,000; set this to your real volume because savings percentages only matter at scale.
7. Leave Strict compatibility on for a safe switch, or turn it off if you are happy to trade a little capability for a bigger saving.
8. If your model is discontinued, turn on Current model is deprecated so replacements that are not cheaper are still shown.
9. Press Get Recommendations. The button shows an analyzing spinner while the catalog is scanned.
10. Read the hero card for the headline saving, then the alternatives table for rates and monthly totals, then the bar chart for a visual comparison.
11. Press Use this on any row to make that model your current one and rerun, chaining upgrades until the savings stop.

Limits and Rules

• Recommendation requests are rate limited to 120 per minute.
• At least one token count must be greater than zero, or the server rejects the request.
• Token counts are clamped to a maximum of 50,000,000 per field, and requests per month to a range of 1 to 100,000,000.
• Only the top three suggestions are displayed, but the counter shows the total number of alternatives found so you know how broad the result set is.
• Discontinued models are never suggested, and models without published prices cannot be compared, so they are skipped.
• With strict compatibility on, a suggestion must match reasoning support, must support tool calls if yours does, and must have a context window at least as large as your requirement.
• Candidates that would cost more are hidden unless you enabled the deprecated switch for a discontinued current model.
• Prices are USD per 1 million tokens from the models.dev catalog, refreshed every 24 hours. List prices only: batch and volume discounts are not modelled.
• No account, API key or captcha is required, and nothing is sent to a model provider.

When to Use It — and When Not To

Use it when your invoice is growing, when you are standardizing a team on one model, when a provider announced a price change, when you inherited an integration still pointing at an old model, or before you commit to a model for a long lived product. It is also useful after a model is discontinued, to see the supported path forward even when it is not cheaper.

Do not use it as the only input. The advisor compares price and a few structural flags. It does not know your quality bar, your latency requirements, your regional data residency rules, your rate limits or your negotiated discounts. A cheaper model that produces weaker answers can cost more in retries and human review. Validate the shortlist with your own evaluation set before switching production traffic, and keep a rollback path ready.

Frequently Asked Questions