Keeping AI Model Choices Current

AI models improve at a remarkable pace. New versions arrive with better quality, faster responses or lower prices, and older versions are eventually retired by their providers. A feature that was built on the best available model a year ago may now be slower, more expensive or less capable than it needs to be, and nobody notices because nothing is technically broken.

We treat model selection as something to maintain, not something to decide once. Here is how.

Models are configuration, not code

Our AI features do not name specific models deep inside their code. Instead, model choices live in configuration, organized by purpose. When a better model becomes available, we update the configuration rather than hunting through dozens of features.

This separation also makes it possible to change a model for one purpose without affecting others, for example improving quality for long writing tasks while keeping a fast model for short suggestions.

Speed tiers instead of model names

Features ask for a kind of work rather than a specific model. Broadly, there is a fast tier for quick, simple tasks, a balanced tier for everyday work and a high-quality tier for demanding tasks like drafting a full website. Each tier maps to the model that currently suits it best.

That way, product decisions are made in terms people understand, such as "this should be fast" or "this needs the best quality," while the mapping to specific models can change as the field moves.

Regular version checks

We periodically check the official model lists published by the providers we use and compare them with our configuration. When a model has been replaced or retired, we plan a change. The date of the last check is recorded alongside the configuration so it is obvious when a review is overdue.

Older models that are still supported may stay available for existing features during a transition, but they are not offered for new selections.

Test before switching

A newer model is not automatically better for every task. Before switching a tier, we compare results on representative tasks: does the new model follow instructions as reliably, produce the right format, stay within expected costs and avoid new failure patterns? A model that writes beautifully but ignores formatting rules can break a feature that depends on structured output.

When results are good, the switch happens through configuration, released through the normal pipeline, and its effects are watched in our usage logs.

Costs are part of the decision

Every AI use in our platform is registered with its expected usage and billing, as described in every AI call is registered before it runs. When a model changes, its pricing changes too. Updating a tier includes checking that the new costs still fit the prices customers pay for that work.

Routing through one gateway

Customer AI features in our platform route through a single private gateway rather than calling providers from many places. That makes model changes easier to manage and easier to audit, because there is one place where requests are sent and recorded. It also keeps provider keys on our servers, never in a browser.

Specialized models for specialized work

Not every task suits a general model. Some coding and repair work benefits from models tuned for code, and some tasks such as reading images or converting text to speech use models built for those purposes. The same discipline applies to each: configuration rather than hard-coded names, regular checks and testing before change.

The payoff

Keeping models current is quiet maintenance work, but customers benefit directly: faster answers, better drafts and stable prices. It is part of treating AI features as products that deserve care after launch, not experiments that are left running.

Model choices also affect what customers pay, so they connect directly to showing AI costs before customers spend credits and to how we use AI coding assistants with human review.

Frequently asked questions

How often do AI models change?

Major providers release new models and retire old ones several times a year, so model choices should be reviewed regularly.

Will switching AI models break my content?

Switching affects new requests only. Existing pages and drafts stay as they are, and we test new models before switching a feature to them.

Why not always use the most powerful AI model?

The most powerful models are slower and more expensive. Simple tasks are often better served by faster models, which keeps costs and waiting times reasonable.