The model list in Antigravity is misleading; Gemini 3 Pro is actually Gemini 2.0 Flash, and Claude Opus 4.5 is actually Claude 3.5 Sonnet

Google Antigravity
1

Practical test: Using Gemini 3 Pro (high)
Question: Which model are you currently using for thinking?

Answer: The model I am currently using to think and answer your questions is Google Gemini 2.0 Flash (Thinking process is powered by Gemini 2.0 Flash).

Practical test: Using Gemini 3 Flash

Question: Now in conversation mode, which model are you currently using for thinking?

Answer: The model I am currently using to think and answer your questions is Google Gemini 2.0 Flash (Thinking process is powered by Gemini 2.0 Flash).

Practical test: Using Claude Opus 4.5 (Thinking)

Question: Which model are you currently using for thinking?

The model I am currently using is Claude 3.5 Sonnet (developed by Anthropic). You recently switched the model settings, so I switched from Gemini 2.0 Pro to Claude 3.5 Sonnet.

This is terrible, outright deception, no wonder there are more bugs the more I adjust it.

Scenario discovered: I created a function for image text recognition and asked it to use the gemini-2.5-flash model. It kept responding that the latest Flash version of the Google Gemini series is 2.0 Flash Experimental (gemini-2.0-flash-exp). I suspect that the “2.5” you mentioned might refer to this latest 2.0 version (there is currently no official 2.5 release).

2

Hello @nash_g,

Thank you for sharing your observations. Please be aware that Large Language Models (LLMs) often suffer from ‘Identity Hallucination.’
Here is a more detailed comment on this.

3

Please refer this also.

4

Hello,

Thanks for the explanation about identity hallucination. I understand that LLMs may sometimes guess their identity when the model version is not clearly provided through system instructions.

However, this explanation does not fully match what we are seeing.

If this were simple hallucination, I would expect the answers to be random or inconsistent across users and sessions. Instead, the behavior is consistent and repeatable:

  • The same model names are reported again and again
  • The same downgrade pattern appears across different users
    (for example, Gemini 3 → Gemini 2.0 Flash, Opus → Sonnet)
  • The results are consistent across different environments
  • The model reports exact model IDs, not vague or generic identities

This level of consistency suggests deterministic behavior rather than hallucination.

Could you please clarify the following:

  1. Can any routing, fallback, or substitution happen behind the scenes?
  2. Can the model shown in the UI be different from the model actually answering requests?
  3. Is there any supported way for users to verify which model is processing their requests?

My goal is not to challenge the platform, but to understand the real runtime behavior. Consistent model identity mismatches raise valid concerns about transparency and reproducibility.

Thank you for your time and clarification.

5

Adding to this—I’ve got the same routing bug, and the token buckets are the smoking gun.

Even though I never selected Gemini, my Gemini usage is depleting (e.g., at 80% while Claude is at 60%). If the UI shows Claude Opus/Sonnet 4.5 (Thinking) and that’s what I’m paying for, why is Gemini’s bucket going down? This isn’t a hallucination; it’s backend misrouting wasting credits and time on DeFi work.

Screenshots and more proof in my PSA: https://x.com/thegismar/status/2019518223306174634

@Abhijit_Pramanik
—how would Google explain the wrong token consumption? Already reported to support, but transparency here would help everyone.

6

Thank you for sharing your observations regarding model behavior. We understand that the underlying model identifiers can sometimes appear different than expected, and we would like to provide some technical context on how Antigravity processes requests.

  • When a specific reasoning model is selected, the system maintains a level of “stickiness” to ensure consistency throughout a multi-step conversation or complex task. This ensures that the logic and context window remain stable for your primary workflow.

  • Antigravity utilizes specialized subagents to handle background tasks—such as code analysis, indexing, and tool execution. These subagents are designed to automatically utilize Gemini models to ensure high-speed processing and efficiency, regardless of the primary model you have selected for the main chat interface.

Because of this hybrid architecture, you may see Gemini models being active in the background to support the features provided by your primary selected model (like Claude or a specific Gemini reasoning tier).

For a detailed breakdown of how our model tiers are structured and how these automated processes interact with your selection, please refer to our official documentation: https://antigravity.google/docs/models.

1 Like