Models Will Keep Changing. Your Workflow Shouldn't.

The best model changes constantly. A few months ago one provider may have been the obvious choice for reasoning, another for speed, another for image generation, and another for cost. Then a new release lands, the rankings move again, pricing changes, an old model gets deprecated, and suddenly the thing you carefully built around one API is already aging. That is normal now. The model layer is moving too quickly to treat any individual model as permanent infrastructure. What should not be normal is rebuilding the product every time it happens. ## The model layer will keep moving That is the idea behind how models work inside Edgaze Workflow Studio. Creators can use current models from OpenAI, Anthropic, Google, DeepSeek, and Kimi, choose different models for different steps, and mix providers inside the same workflow. One step might need strong reasoning, another might need something cheap and fast, and another may need a completely different modality. There is no reason the entire workflow should inherit the strengths, weaknesses, pricing, and limits of a single provider just because that was the API you happened to integrate first. More importantly, switching models should be boring. If a better model comes out, the creator should be able to change that step without rewriting the surrounding workflow around a new SDK, request format, authentication system, response structure, or provider-specific set of quirks. The workflow logic stays where it is. The model underneath it changes. That sounds like a small abstraction until you have maintained AI software for more than five minutes. Provider APIs are not identical. Model names change. Parameters differ. Capabilities move around. Models get retired. Some providers keep older versions around for a long time, while others eventually tell you that the model your production system depends on is going away and wish you the best of luck. ## Choose models without coupling the workflow Edgaze tries to absorb as much of that mess as possible at the runtime layer. If a provider deprecates a model, Edgaze can detect that the workflow is pointing at something no longer supported, map it to the appropriate replacement, and continue the run rather than turning a provider lifecycle change into a broken marketplace workflow. Creators should still be able to deliberately choose and optimize models, but they should not have to patrol every workflow just to make sure yesterday's model identifier still exists. There is an important distinction here between **choosing the model** and **being coupled to the model**. Creators should absolutely be able to choose. Models behave differently, and pretending they are interchangeable would be nonsense. Some are better at specific tasks, some are much cheaper, some are faster, and some have capabilities the others do not. Those differences matter when building a good workflow. What should disappear is the infrastructure tax around those choices. A workflow should describe the job being done. Call a model here. Extract structured data there. Branch based on the result. Pass something into another model. Combine the outputs. Return the outcome. The fact that one of those calls currently happens to go through OpenAI and another through Google should not dictate the architecture of everything around them. ## Buyers should not have to manage the model layer This also matters for buyers because they should see even less of the model layer than creators do. Someone running a workflow on Edgaze should not need five API accounts, five sets of credentials, five billing relationships, and a spreadsheet explaining why three providers charged them this month. They should not need to know which endpoint an individual step uses or whether a model was renamed last Tuesday. They should care about whether the workflow produces the result they paid for. That is the level we want the abstraction to reach. Creators can still go deep when model selection matters. Buyers should barely have to know the abstraction exists. There is always a temptation in AI products to build the product around whichever model currently looks unbeatable. That works right up until it is not unbeatable anymore, which lately can happen before the launch post has finished circulating. ## A catalog that changes without breaking workflows The safer assumption is that the model layer will continue to move aggressively. New providers will become competitive. Current leaders will leapfrog each other. Prices will fall. Capabilities will expand. Model families will disappear. So we are designing around that instability rather than pretending it will stop. The current catalog is the same one Workflow Studio and billing use. The names below stay in sync automatically as the catalog changes. Prices, hosted rates, and migration rules live on the [Model Catalog & Pricing](/docs/model-catalog-pricing) page. ## Chat models ```docmodels {"modality":"chat","namesOnly":true} ``` ## Image models ```docmodels {"modality":"image","namesOnly":true} ``` ## Embedding models ```docmodels {"modality":"embedding","namesOnly":true} ``` Models can change. The workflow should keep running.