Catena Labs

Augmenting Claude 3.5 Sonnet: Introducing Super Sonnet Mix

Joao Fiadeiro

At Catena Labs, we use Claude 3.5 Sonnet every day for many tasks — from content creation to coding. It's become an essential part of our daily workflows. However, as we scaled our usage, we started hitting some practical limitations that impacted our productivity.

One frustrating issue was hitting rate limits at crucial moments. We also found ourselves wanting built-in capabilities that other platforms offered, like extending the model’s knowledge cutoff with web search and adding image generation capability. Managing multiple provider APIs to work around these limitations added unnecessary complexity to our development process.

To solve these challenges for ourselves, we built Super Sonnet Mix, a recipe for Sonnet that we're now making available to others who may find it useful.

super-sonnet-mix

Here's how it can help your team

First, it mitigates rate limit headaches. When you hit limits with the primary Anthropic API, the mix automatically fails over to AWS Bedrock or Google Vertex. You get higher effective rate limits without managing multiple API keys or handling provider-specific errors.

Note that these failover endpoints support all the expected Anthropic API features (for example, prompt-caching).

Second, it adds web search capabilities using Exa and Tavily. This means no more knowledge cutoff limitations—Claude can now access and reason about current information. Instead of "I apologize, but my knowledge is cut off from April 2024," you get accurate, up-to-date responses.

Third, it integrates image generation with Flux (Schnell). Rather than Claude's usual "I cannot generate images" response, you can now create visuals directly through the same API. This makes the chat experience more complete and reduces the need to context switch between different tools.

Integration is straightforward. If you're already using an SDK, just update two things:

  • Set the base URL to https://api.crosshatch.app (and append /v1/ if you are using the OpenAI API instead of the Anthropic API)
  • Change the model to super-sonnet-3.5-20241022
  • In your request, add the tools you want to use. For example, "tools": [{ "name": "crosshatch:search" }]

Your existing code will work as before, but with the enhanced capabilities and improved reliability.

We've been using this internally for several months now, and it's made a significant difference in our development workflow. If you've encountered similar challenges with Sonnet, we think you'll find it useful too.

Augmented models like this one can help you build your apps faster and more reliably. In the future, we may build more Super Mixes that supercharge upcoming models that may be missing a thing or two.

Looking ahead, we're excited about what 2025 holds as we explore the future of agentic workflows, identity, and payments. This Super Sonnet Mix has been playing a critical role in that work. We’re excited to share more about that work soon!

You can try out Super Sonnet Mix at crosshatch.app. Join our Discord community if you'd like to discuss your experience or share feedback.