Skip to main content
For the complete documentation index for agents and LLMs, see llms.txt.

Compare ModelsBeta

Beta
This is coming — this page describes a feature that's still being rolled out and may not be available to you yet.

Use the Models page to compare quality and price across the models available on the platform. This helps you pick a model for your pipeline, for example the cheapest model in a class or the best model from one provider.


About This Task​

Models are rated by class, quality, reasoning, and price. The classes are Top tier, Budget, Balanced, and Unknown (no benchmark data). Each model also gets a value verdict: Premium, Great value, or Fair. To learn what the classes, ratings, and value verdicts mean, and where the data comes from, see Model Classes.

To list models programmatically, including embedders and rankers, see List Embedders and Rankers with the REST API.

Filter, Group, and Sort Models​

  1. Open the Models page.
  2. To filter by class, use the Class filter. You can select several classes.
  3. To narrow the list to the models with the best value verdict, use the Best value filter.
  4. To group models, choose Provider, Tier, or None in Group by. Tier groups models by class.
  5. To change the order, sort by family (newest first), best value, highest quality, or lowest cost.

Check a Model's Ratings​

Hover over or focus a model card to see its class, Quality, Reasoning, and Price ratings, and an insight such as similar quality to another model at a third of the price. The summary also shows the raw LMArena score and rank, so you can check the ratings against the source.

Compare Quality and Price in the Chart​

  1. On the Models page, in Display as, choose Chart. Models toward the top are higher quality, and models toward the left are cheaper per 1,000 typical requests. Cost increases from left to right.
  2. Choose how to group the models. The chart colors the models by the grouping.
  3. Click a point to open the test drawer for that model.

Models without benchmark data can't be plotted. To see them, click the chart's link to the cards, which opens the Unknown section in the list view. To go back to the cards at any time, choose List in Display as.

Test, Build, and Compare Models​

  • Use Test on a card to try the model.
  • Use Build to start a pipeline from the model.
  • Use Compare in the list view to compare models side by side. Compare isn't available in the chart.

Some Amazon Bedrock models are available in both a global and an EU variant. They appear as one card with a Global | EU switch, and Test, Build, and Compare act on the variant you pick. A model that is offered only in the EU shows an EU badge.

Unknown models have no quality comparison, but you can still test them and build from them.

Try a Model in the Test Drawer​

Use the test drawer to chat with a model before you build anything and check its answers, tone, and speed against your use case. Nothing you send is stored.

  1. Open the test drawer for a model, for example with Test on a card.
  2. Optionally, expand System prompt to set the instructions the model follows. The prompt is collapsed by default. The system prompt you set here is carried into the pipeline you create from the drawer.
  3. Optionally, click the Parameters icon in the drawer header to tune the model's parameters.
  4. Type a message in the input bar and press Enter to send it. Press Shift+Enter to start a new line.
  5. Read the reply. The line below the answer shows how long the model took and how many tokens it used, so you can compare models on speed and cost.
  6. To start over, click the Clear chat icon in the drawer header. It's available once the conversation has a message or an error.

When you're happy with the model, click Start building with this model in the drawer footer. This creates a pipeline that starts with the system prompt and parameters you set in the drawer.

If the model doesn't respond within 60 seconds, the drawer stops waiting and shows a message. Try again, or try a different model. If the model returns an error, the error appears in the thread.