Gemini 3.0 Pro Leaked: Some 2.5 Prompts Accidentally Routed to New Model
3 min read
by Ufuk OzenGemini 3.0 Pro
Gemini 2.5
AI Leak
Google AI
Riftrunner
A routing error accidentally reveals Gemini 3.0 Pro, with some Gemini 2.5 prompts being sent to the unreleased model. Here's what we know about the leak, the 'Riftrunner' checkpoint, and how to potentially access it.
Gemini 3.0 Pro Leaked Through Accidental Routing
The AI world is buzzing with a new leak, and it's a big one. It appears that some prompts sent to Google's Gemini 2.5 Pro are being accidentally routed to the unreleased Gemini 3.0 Pro model. This surprising development has given the community a sneak peek at the next generation of Google's AI.
This is what one user one-shotted with just the word "aesthetic":
BREAKING: Gemini 3.0 Pro has been spotted in Gemini Enterprise in the Agent model selector.Gemini 3.0 Pro Preview
How Is This Happening?
According to reports, the leak is due to a router that sometimes directs prompts from the public-facing Gemini 2.5 Pro to an internal Gemini 3.0 Pro model. This seems to be happening most frequently when using the Gemini App on mobile in "Canvas mode".
While not every prompt is being rerouted, enough users have experienced it to confirm the existence and power of the upcoming model.
How To Potentially Access Gemini 3.0 Pro
While there's no guaranteed method, the community has found a series of steps that seem to increase the chances of getting a response from Gemini 3.0 Pro:
- Download the Gemini App on your mobile device.
- Sign into your Google account.
- Select Canvas mode.
- Choose the 2.5 Pro model.
- Enter your prompt and wait.
It's important to remember that this is not an official release, and results may be mixed. The routing seems to be inconsistent, but many users have reported success.
"Riftrunner" and KingBench Performance
The specific version of Gemini 3.0 Pro that has been accessed is codenamed "Riftrunner". Early benchmarks on KingBench show impressive, if varied, results.
According to one source, this is the "worst CKPT yet of all Gemini-3 CKPTS." However, that same source clarifies:
However, it's still way ahead of all current gen models.Beats Sonnet 4.5 by 15%, 14% worse than X58 (Best CKPT yet).
This indicates that even the lower-performing checkpoints of Gemini 3.0 are a significant leap over the current state-of-the-art models from competitors.
What This Means for the AI Landscape
This accidental leak has several major implications:
- Confirmation of Gemini 3.0 Pro: We now have concrete evidence that Gemini 3.0 Pro is not just in development, but actively being tested internally.
- Significant Performance Gains: The "Riftrunner" benchmarks, even if from a "worse" checkpoint, suggest a substantial jump in capability.
- Competitive Pressure: This leak puts pressure on OpenAI, Anthropic, and other AI labs, as it signals that Google is on the verge of another major release.
While we wait for an official announcement from Google, the AI community will be eagerly experimenting and trying to get another glimpse of the future. Have fun, everyone!
