Who Powers Betty?

As many of you already know, we don’t train our own language models for Betty to run on. Instead, we leverage existing models and providers – swapping them out as needed to keep up with the level of intelligence required and to improve performance over time as AI evolves. 

Up until now, we’ve relied on the major players in this space: OpenAI, Google, Microsoft, and Anthropic. They’ve been the frontier developers – well-known, predictable, and for a long time, the only options smart enough to make sure Betty could follow all the nuanced instructions she needs for the variety of tasks thrown her way. 

But they’re no longer the only ones in the game. Other providers are catching up fast – offering models that can handle much of what we need, often cheaper and significantly faster. Using some of these models at the right points could not only make Betty more responsive but also open the door to new use cases we’ve had waiting in the wings. 

Since we’ve been testing a few of these options, I wanted to share what we pay attention to when evaluating providers, how we make those decisions, and where Betty’s customers – and the broader association community – fit into the process. 

What We Look For

When choosing both models and providers, there are several key factors that each have the power to make or break an option: 

• Capability

• Performance

• Capacity
• Security 

Capability

This one’s simple: can the model actually do what we need it to do? 

AI has changed a lot since Betty was first built. Early on, we had to get very good at squeezing great results from what are now considered lower-end models, but as needs have evolved, so have expectations. Some of the layered, nuanced instructions Betty now handles would’ve stumped even the best models from a couple years ago, so we rely on the improved intelligence from how AI has advanced. 

While open-source models have come a long way, most still weren’t quite up to par for our needs – until recently. That’s starting to change, and it’s why we’re exploring them more seriously now. 

Performance

Performance is all about speed. 

Some of the smartest models today can reason through problems, handle complex data, and even do math reliably – but that kind of reasoning comes at a cost: time. Betty already has multiple steps to complete for each request, and if every step added just a few extra seconds, that would add up fast. Smaller, faster models are nice, but the performance of the exact same model can vary widely depending on the provider as well. 

For most tasks, we can’t afford that tradeoff. So, while we’re excited by advanced reasoning, we have to balance intelligence with responsiveness. 

Capacity

Capacity is the question of whether a provider can handle the volume Betty needs to stay smooth and fast. 

This used to be a concern even with the big names, but they’ve since scaled up. You might hear this described in terms of tokens or requests per minute/hour/day, and those limits can be a serious bottleneck with smaller providers. 

If there’s any chance we’ll hit rate limits during traffic spikes, we simply can’t rely on that provider – no matter how impressive their models are. 

Security

This one is non-negotiable. 

Security is about how your data is handled as it passes through these providers. It’s a mix of clear terms of use and trust in how the vendor enforces them. 

For example, OpenAI originally lost some trust by being unclear about data use when ChatGPT was first released, but they’ve since clarified those terms. When using their API (like we do with Betty), your data isn’t used for training. 

Other providers, especially those who just host models rather than train them, may actually have less incentive to retain or use your data – which can be a good thing. But every provider needs to prove that before we consider them. 

Sometimes, people will confuse trusting models with trusting providers – there are some cases where a provider may not be one you would like to use, but the models they’ve created are perfectly fine when hosted elsewhere. 

Why This Matters Now

This has been top of mind lately because we’re making a big push this quarter to dramatically improve Betty’s performance – both how quickly she starts to respond and how fast the responses themselves come through. 

There are models and providers out there that blow current options away in terms of speed, and a few features we’ve been holding back might finally become viable with those improvements. The catch is making sure those providers also meet our requirements on capability, capacity, and security before we make any move. 

And just to be clear – if/when we do want make a switch, it won’t happen quietly. We’ll always communicate ahead of time so you can review the terms, confirm the security standards meet your needs, and opt out if preferred. 

Looking Ahead

The major providers are stable and familiar, which makes them a safe bet. But if you don’t need frontier-level reasoning, the ecosystem of smaller, faster, and cheaper options has become really compelling. 

Our job is to stay ahead of those developments – to keep Betty getting smarter, faster, and more adaptable – without compromising on trust or security. That balance is what keeps her reliable for our customers and for the wider association space.