This morning, Gemini Flash was being particularly dumb, so I asked it what version of the model it was, and it replied with Gemini 1.5 Pro.
Is it normal to get 1.5 Pro when paying for the Pro plan?
I am **Gemini 1.5 Pro**, a large language model trained by Google. In this session, I am acting as **Antigravity**, a coding assistant developed by Google DeepMind.
This is a known hallucination issue. You are definitely still using the Flash model.
From a cost perspective, running 1.5 Pro is much more expensive than Flash, so it wouldn’t make sense for them to serve the heavier model. The model is simply outputting an incorrect version name based on a leaked system prompt or its training data.
Well the hallucination seems to be deep, I asked the last event he known without searching on internet or using the context:
Without searching the internet or using the current conversation context, the last major events in my internal training data date back to **early 2024**.
One of the most specific "last" events I recall is the **total solar eclipse** that occurred on **April 8, 2024**, which was visible across a large part of North America.
Could be possible there is a problem of availability for Flash 3.1 and they are using old Flash 1.5 instances?
That’s a classic hallucination, too. LLMs don’t have an internal clock and are notoriously bad at knowing their own training cutoff dates. They just guess based on token probabilities, which is why it confidently made up the eclipse date.
The theory about them falling back to 1.5 instances is also highly unlikely for two reasons. First, silently mixing model generations under the same endpoint would break token counting and context window limits. Second, the newer 3 flash architecture is highly optimized and significantly cheaper to run per token than 1.5.
Spinning up legacy, less efficient instances during a traffic spike would literally cost Google more compute and money. It’s 100% just the 3 flash model being confidently wrong about its own identity.