TAU-HOME.COM
LOADING

Google Next-Gen 'Gemini 4 Argon' Spotted in Covert A/B Testing Under Gemini 3.8 Flash on Antigravity

TestingCatalog reports that Google is quietly A/B testing its unreleased model, Gemini 4 Argon, under the Gemini 3.8 Flash label on Antigravity. Early outputs r

tau · October 13, 2026

#Google #Gemini #Gemini-4-Argon #Antigravity #AI-Models #AB-Testing

Google Next-Gen 'Gemini 4 Argon' Spotted in Covert A/B Testing Under Gemini 3.8 Flash on Antigravity

On October 12, 2026, AI intelligence tracker TestingCatalog (@testingcatalog) reported that Google has reportedly been A/B testing an unreleased next-generation foundation model, 'Gemini 4 Argon', under the label of 'Gemini 3.8 Flash' on the Antigravity developer platform.

Conceptual diagram of Google's Gemini 4 Argon covert A/B testing under Gemini 3.8 Flash on Antigravity with world generation comparison

Image source: @testingcatalog

The observation indicates that Google may be testing candidate weights against live user traffic ahead of an official launch, masking the model behind an existing Flash tier to evaluate output quality. The findings align with observations from developers who noted sudden, unexpected quality jumps during routine Gemini 3.8 Flash sessions.

A/B Testing Behind the Gemini 3.8 Flash Label

According to TestingCatalog, Google has been conducting A/B testing by routing select sessions on Antigravity invoking the Gemini 3.8 Flash endpoint to the unreleased Gemini 4 Argon backend.

  • Observed Scope: Beyond Antigravity, community members also reported occasionally observing similar testing under Gemini 3.8 Flash on the web.
  • Traffic Routing: Rather than introducing new model selectors, the testing appears to use an A/B routing split while retaining the existing 3.8 Flash identifier.
  • Explaining Quality Spikes: Users had recently reported sudden quality jumps during routine Flash sessions; community observers noted that this A/B testing report would explain the unexpected spikes in output quality previously experienced by users.

Silent A/B splits to evaluate user interaction in production before public release remain a common practice across frontier AI labs.

World Generation Comparisons and Early Evaluations

The centerpiece of TestingCatalog's report is a side-by-side comparison using a 'World gen prompt' to evaluate the current Gemini 3.8 Flash against the candidate Gemini 4 Argon model.

  • Detail and Quality Leaps: In the published side-by-side comparison videos, the suspected Gemini 4 Argon demonstrated a pronounced qualitative leap over the existing Gemini 3.8 Flash model, delivering substantially higher detail density and overall generation quality.
  • Positioning Against GPT-6.1: In a follow-up comment, TestingCatalog stated that Gemini 4 Argon was slightly better overall than GPT-6.1, while performing significantly better in select areas.

Additional sample outputs and early comparative analyses were published directly on TestingCatalog's website (testingcatalog.com).

Community Feedback and Release Caveats

While the sighting of Gemini 4 Argon generated interest across the community, user feedback also highlighted key caveats ahead of any official release.

  1. Domain-Specific Feedback: While the world generation output was widely praised, user feedback indicated that the model is still not that good at coding (@mrfanduu), reflecting mixed impressions depending on the task domain.
  2. Benchmark Contamination Concerns: Community observers pointed out that serving unreleased candidate models under an existing identifier during A/B testing could inadvertently cause benchmark contamination by skewing evaluation data associated with 3.8 Flash.
  3. Pre-Release Status and Unconfirmed Specifications: Google has not officially announced Gemini 4 Argon, nor has it confirmed release schedules, final branding, or formal specifications. While TestingCatalog responded that a release would be "soon," official specifications await Google's formal disclosure.

Sources