A Flagship Hiding Behind an Old Version Number
In mid September 2026 several X users and Chinese tech outlets reported that Google had submitted an unreleased model to Arena, the third party leaderboard where vendors can enter anonymous candidates. The model reportedly served under the label gemini-3.8-flash while carrying the internal codename Argon, and was described as the next flagship Pro tier rather than another Flash iteration. Google did not comment.
Why Enter a Blind Test Under a Disguise
Arena lets users vote head to head without knowing which system produced each answer, which turns the leaderboard into a stream of unbiased preference data. Using an old version string instead of full anonymity is a deliberate middle path. It leaves a trail back to a real product line for internal traceability, while denying competitors a public signal about where the next capability ceiling sits before launch. The same mechanism also lets a team read genuine Elo movement without anchoring outside expectations to a brand name.
Early Hands On Reports Focus on Visual Structure
The tests circulating online leaned heavily on vector generation, starting with the now standard stress case of an SVG of a pelican riding a bicycle. SVG is an unforgiving probe because it demands coherent paths, geometry and layering rather than pixel soup. Screenshots shared by testers described the result as unusually clean and were posted alongside a same prompt comparison against a top OpenAI model.
Longer builds drew similar attention. One developer reported a scroll driven sketch landing page, where pencil and graphite strokes darken as the reader moves down the viewport, completed in about fourteen minutes. On the 3D side, reports described an Airbus H145 helicopter model with five switchable liveries, cockpit and tail rotor detail, and a wireframe view, built in roughly ten minutes, plus a voxel pagoda offering four lighting modes and six camera positions with floating lantern effects at night. Game style outputs were claimed at similar speed, including a block building sandbox and a kart racing prototype.
Taken together, these anecdotes point at a single underlying capability: holding structural, spatial and code relationships stable long enough to finish something interactive.
Leaked Specifications and Benchmark Tables
Reported backend limits describe a large jump from the current generation. Tester accounts claimed a ten million token input ceiling, a 256,000 token output ceiling, persistent memory across sessions, and network access without routing through the API layer. The output limit alone would be four times the previous 64,000 token ceiling, and a two million token context window has been rumored separately.
A benchmark image spread at the same time, and if the numbers hold, the model leads on three measures: long horizon software engineering at 88.7 percent against 86.9 percent for GPT-6 Astra, multidisciplinary reasoning at 72.1 percent against 67.3 percent, and desktop automation at 86.8 percent, ahead of both Astra and Fable. None of these figures had independent verification at the time of reporting, and pre release scores routinely shift before a general availability build.
Pricing chatter placed the model as the cheapest of the three frontier tier systems then discussed, at a reported $2.25 per million input tokens and $11.25 per million output tokens.
From a Cancelled 3.5 Pro to a Skipped Generation
The leak landed after an unusual stretch for Google's flagship line. The last Pro release dated to Gemini 3.1 in February, and the widely expected 3.5 Pro, announced at I/O, reportedly never shipped because internal evaluations judged its agent capability short of target. Google then shipped 3.6, 3.7 and 3.8 as Flash models, which serve volume rather than the capability headline. Skipping straight to a fourth generation is consistent with a reset rather than an increment.
Recursive Self Improvement as the Real Claim
The most speculative element concerns training method. Google DeepMind's chief strategy officer Jasjeet Sekhon spoke at a Berkeley event about recursive self improvement as a load bearing part of the investment case, framing it as a possible accelerant beyond exponential capability growth. Around September 14 a DeepMind paper on Dream RSI became public, describing agents that refine their own search strategy across exploration rounds. Connecting that research to a specific product is inference, not confirmation, but the direction of the published work is not in dispute.
What a Rumor of This Scale Actually Signals
Pre release builds usually improve further during final optimization, and some analysts argued the shipped version could sit materially above what leaked. The more durable reading is competitive: a lab entering blind tests under a stale label is collecting preference data while holding back a public narrative, and the leaked strengths in SVG, 3D and interactive generation target a different buyer than chat quality benchmarks do. Confirmation still requires an official release, published system cards, and third party reruns of the claimed numbers.
