The stealth deployment began on September 17, when a model calling itself “Gemini 3.8 Flash” surfaced on the LMArena leaderboard. Testers observed the system generating SVG graphics and interactive 3D outputs capabilities that far exceed standard budget-tier models. By September 26, a fuller checkpoint operating under the internal codename Argon surfaced with attached benchmark scores.
Record-Breaking Scores and Operating Capability
The leaked benchmarks include 88% on DeepSWE v1.1 for agentic coding, 95.3% on Terminal-bench 2.1, and 86.8% on OSWorld-2.0, a metric that evaluates an AI’s ability to operate a computer independently. Developers who analyzed the data noted that these scores position the unannounced model above OpenAI’s GPT-6 Astra and Anthropic’s Claude Fable 5.1 across reasoning, coding, and complex agentic tasks.
Google has not confirmed the checkpoint’s identity, performance figures, or an official release date. The public release of the actual Gemini 3.8 Flash occurred on September 2, providing convenient cover; testers grading responses blind could not easily distinguish between the current budget model and next quarter’s flagship.
Pricing Pressures and the Fall AI Race
Beyond performance metrics, the leaked table details projected pricing of $2.25 per million input tokens and $11.25 per million output tokens. For comparison, OpenAI’s GPT-6 Sol runs at $2 and $10, while Anthropic’s frontier pricing sits higher. If confirmed at launch, the figures suggest Google intends to undercut competitors while delivering superior performance.
Timing complicates the competitive landscape. Anthropic released Claude Fable 5.1 on September 1, followed by Claude Opus 5.5 on September 22. OpenAI countered with GPT-6 Sol and GPT-6 Luna, driving down input costs. The arrival of a high-performance Google challenger disrupts a market currently locked in an intense capability and pricing war.
Skepticism and Verification Challenges
Industry analysts maintain caution regarding blind benchmark data. Similar leaks, such as previous claims surrounding Claude Sonnet 5.5 and GPT-6 Sol, required rigorous methodological scrutiny. While standardized tests measure specific competencies like terminal operation, they often leave open questions about how models handle ambiguous, real-world enterprise workflows.