We decided to publish a general note about the AI models we work with when we need intelligence systems within the agency, along with an announcement about a short outage in our systems. If you too have placed an intelligence system at the center of your systems, the points we raise here are worth watching.
This piece is written within the scope of AI optimization (AIO).
For Ads, for SEO and for our digital inferences, we work above all with Gemini as our AI system. We use different models depending on the tool we have built, and the problem we ran into sat on exactly that axis.
Our Turksemveri system generally uses the latest version of Gemini. We need reasoning at the highest level for the logical analysis of the content on our SEO clients’ websites.
For the free tools across our site we generally choose the Gemini model that sits best on the price and performance balance. The main reason is that we supply all the data, we state how it should think and we state how it should build the best sentence. Reasoning in these tools is minimal, and the basic element is the balance of speed and correct expression. Recently some Türk SEM tools that need an intelligence engine were down for a few weeks. The reason was that we had not been told a Gemini model was being retired.
Gemini 3.1 Pro scored 77.1% on the abstract reasoning (ARC-AGI-2) tests in particular, doubling the performance of the earlier models.
So that you do not run into a similar problem in your own work, below I will share the Model Aliases codes; with them you can go on serving without interruption.
What Are Model Aliases?
Technology moves so fast in the world of AI that the “smartest” model you use this week can hand over to a newer version next week. Model aliases are a technical “routing” mechanism that lets you always use the most current and stable AI version without changing your software code.
Normally, when you connect to an AI model, you use a very specific version code such as gemini-3-pro-preview-0125. When that model is retired (deprecation), though, your code starts throwing errors. Model aliases step in here and let you work through a fixed “alias” instead.
How Does It Work?
In this system, which acts as a bridge, your software gives the provider (Google or OpenAI, for instance) this instruction:
“Give me the latest and best model in this series.” (whoever thought of this, thank you)
- Instead of a fixed code:
gemini-3-pro-preview(when the model is retired, the system stops.) - Using an alias:
gemini-pro-latest(when the model is updated, the provider routes this name automatically to the new model, for instance Gemini 3.1 Pro.)
Why Does It Matter?
- Continuity: It minimizes the risk of an outage at API level during model updates.
- Zero code changes: When a new intelligence engine appears, you do not have to update thousands of lines on the backend.
- Automated performance: It keeps the system running on the current architecture with the highest reasoning and speed capacity.
In short, model aliases are like a gearbox that makes your AI infrastructure future-proof and shifts your “intelligence engine” up a gear automatically in a dynamic world.
Model Alias Codes
| Provider | Alias | Description / routing |
|---|---|---|
| Google Gemini | gemini-pro-latest | Automatically routes to the most current Gemini Pro model (at present Gemini 3.1 Pro). |
| Google Gemini | gemini-flash-latest | Points to the most current fast Flash model. |
| Mistral AI | mistral-large-latest | Calls the latest version of Mistral’s most powerful “Large” model. |
| Mistral AI | mistral-small-latest | Points to the most current efficiency-led “Small” model. |
| Mistral AI | codestral-latest | Uses the latest version of the model built for software and coding. |
| OpenAI (GPT) | gpt-4o | Routes to the most current stable version of the GPT-4o series. |
| OpenAI (GPT) | gpt-3.5-turbo | The automatically updated, most stable name in the 3.5 series. |
| Anthropic (Claude) | sonnet | Points to the most current Sonnet model (at present 4.6). |
| Anthropic (Claude) | opus | Calls the most current and most capable “Opus” model (at present 4.6). |
| Anthropic (Claude) | best | Automatically selects the most capable model available (usually Opus). |
An Alternative to Model Aliases
You build a list of the intelligence engines that suit your system best. You put the options in order on that list and set up a cron job. At a set hour of the day, that cron job tests the AI models one after another; if a model has been retired it tests the next, then the next again. Whichever model it last worked on becomes the central intelligence engine for that day, the one that signalled it was running. It tries this every day, and every day one engine from your list, one that works, is the one running. That is how dynamic engine selection and failover keep the risk to a minimum.
Summary
The tools that stopped working because of the model outage Türk SEM went through led us to research the Model Aliases codes in detail, and we wrote this piece so it could serve as a guide for every SaaS developer. Your choice of model matters a great deal. Analyze the results of each model separately, and if you use a model code, follow the official sites so you are ready for retirements.
