A checkpoint of Gemini 4 Pro, Google’s next flagship model, showed up on September 26, 2026 on the benchmarking platform LMArena disguised under the name Gemini 3.8 Flash, the Flash model the company had already officially launched on September 2. Pankaj Kumar (@pankajkumar_dev), a developer who tracks large language model releases closely, was the first to flag it, in a thread posted on X.
Google and DeepMind have, however, confirmed nothing. What’s on the table is a pattern that has repeated itself for weeks now, independent testers noticing output that’s too good for a “Flash” model (the lightweight, fast tier of the Gemini family, built for quick and cheap answers), and, according to industry checkpoints ahead of a public launch
What happened on LMArena
LMArena is a blind-testing platform where users compare the answers of two models without knowing which is which, then vote for the better one. Companies also use it to run internal checkpoints under a pseudonym ahead of an official announcement, gathering real preference data without revealing what they’re building.
The first sighting dates back to September 14, when a tester going by Lentils wrote that “Gemini 4 Pro checkpoints started showing up internally a few days ago.” On September 17, another tester, Bee, posted a demo of a one-page site the model had generated, surprised at how clean it looked. On September 19, a third account, Lumina, reported finding the checkpoint on Arena routed under the label gemini-3.7-flash, as reconstructed by NPowerUser. Sources diverge on this point. Other screenshots circulating online, including Pankaj Kumar’s from September 26, show gemini-3.8-flash instead. Neither label matches any name Google has officially used for this release.
Why the name is confused with the Gemini 3.8 Flash already out
This is where things get complicated. The company did release a model called Gemini 3.8 Flash on September 2, 2026, alongside a Cyber variant for cybersecurity. That launch came after weeks of buildup, including the Wall Street Journal’s reporting on the codename Skimaki. The checkpoint now in question is a different animal. It runs under a label that reuses, more or less (given the conflicting gemini-3.7 and gemini-3.8 versions), the already-public name, but the performance testers describe has nothing to do with a budget-tier Flash model. This is a case of pseudonym routing, now a de facto standard technique for testing a Pro-tier model without exposing its real name in a public arena.
What’s the checkpoint’s actual name? The internal codename citing the working name that surfaced through leaks
What testers saw, namely speed and design
Pankaj Kumar describes a model that’s “damn fast, almost twice as fast as other flagship models.” That’s a subjective impression, not a published benchmark per se, but it lines up with what TestingCatalog reported, response times between 2.4 and 20 minutes depending on the task, against the extended reasoning mode that earlier checkpoints almost always pushed to the upper limit.
The point where several independent sources converge is aesthetic taste. Bee, the tester who generated the one-page site, said they were surprised at how “clean and polished” the result was, adding that earlier Gemini models’ design-taste problems finally seem solved. Kumar goes further, writing he’d seen “the best design taste I have seen in any model.” These are human judgment calls, not an objective score, but the fact that different testers, on different prompts, reach the same conclusion is a signal worth reporting.
Gemini 4 Pro New Checkpoint : My Impressions
– A new Gemini 4 Pro checkpoint came to Arena under the name “Gemini 3.8 Flash.”
– I tested it and its a damn fast model, almost 2× faster than other flagship models.
– It has very good design taste, honestly, I would say best taste… pic.twitter.com/0EilqDrY69— Pankaj Kumar (@pankajkumar_dev) September 26, 2026
The extra detail in 3D graphics
On the 3D front, the reported tests are more concrete than taste judgments. Some developers asked the checkpoint to generate a Formula 1 car in Three.js (the JavaScript library for 3D graphics in the browser). The result included a monocoque chassis, double-wishbone suspension with upper and lower arms, carbon-fiber materials with PBR shaders (Physically Based Rendering, the technique that simulates how light behaves on real materials) and adjustable HDR lighting. Another test produced a photorealistic, explorable airship, also in Three.js, generated in under ten minutes with no follow-up code fixes.
The most direct comparison came on September 26, when testers pitted the checkpoint against Claude Opus 5.5 on a physics simulation of an airplane flying over the sea. The verdict reported by NPowerUser is blunt. “Gemini 4 Pro beat Opus 5.5 by a wide margin on this one. Better colors, better reflections, more detail on the plane.” In the water reflections, Opus showed ripples with visible artifacts, while the Gemini checkpoint produced more coherent wave motion. Same gap in a scene with dinosaurs rendered under volumetric lighting, where skin and muscle detail came out sharper in the Gemini version.
When Gemini 4 Pro might arrive
Here we’re in prediction territory, and it should be treated as such. Kumar writes that “if the pace of improvement keeps up like this, they might launch it in early-to-mid October.” That’s a guess based on how quickly the checkpoints have followed one another (September 14, 17, 19, 26), not a date the company has communicated. The reference to “RSI that’s working” in Kumar’s tweet points to Recursive Self-Improvement, the idea that a lab is making its own models improve faster and faster by using those same models inside the development loop, a concept debated within the AI community, used here colloquially rather than as verified data.
Come October 2026, the release calendar looks crowded regardless. DeepMind’s new chief has already said Gemini 4 will arrive as soon as possible, and the Arena checkpoints should be read in that light, intensive testing ahead of a launch Google wants to time to the window where competitors are moving fastest.
Why it matters even if you don’t use Gemini
The direct comparison with Claude Opus 5.5 isn’t an isolated detail. October is shaping up as a month of direct confrontation among the top labs. Anthropic already has Opus 5.5 in the field (which introduced an experimental reset of usage limits), OpenAI is betting on GPT-6 Astra, which in these same weeks is even being tested behind the wheel of a car, and Siraya Technologies is already citing comparisons underway with Claude Fable 5.2 and Grok 4.7. If the Argon checkpoint keeps in production what testers are describing now, Google would have a concrete argument on generation speed, historically one of the Gemini family’s weak points against the competition.
What is Gemini 4 Pro’s Argon checkpoint?
Argon is the internal, Google-unconfirmed name of a Gemini 4 Pro checkpoint spotted on LMArena starting September 14, 2026. It was routed under the label of an already-public Flash model to hide its real identity during blind tests with users.
Frequently Asked Questions (FAQ)
1. Is Gemini 4 Pro already available to the public?
No. What’s circulating are test checkpoints routed on LMArena under decoy names, not a public release. Google has confirmed neither the checkpoint’s existence nor a launch date.
2. What’s the difference between this checkpoint and the Gemini 3.8 Flash already out?
They’re two different models, despite the similar label. Gemini 3.8 Flash is the lightweight model officially launched on September 2, 2026. The checkpoint testers are describing is instead attributed to Gemini 4 Pro, the flagship tier, and is being run on Arena under a name echoing the already-public one.
3. What does a “256k token output limit” mean?
In simple terms, it’s the maximum length of a single response the model can generate in one go, measured in tokens (text units, roughly three quarters of a word in English). The figure reported by TestingCatalog for the Argon checkpoint is four times the previous checkpoint’s, which was capped at 64,000 tokens, that is, more room to generate code or long documents without splitting the answer into multiple passes.
4. Why do companies test models under a fake name on LMArena?
To gather real preference data without revealing what they’re about to launch. Users vote blind between two answers, so the lab finds out how the model performs against the competition before even announcing it, keeping rivals from preparing in advance.
5. Are the October launch predictions official?
No, they’re a guess from whoever tested the checkpoints, based on how often they popped up between September 14 and 26. Google has not communicated any release window for Gemini 4 Pro.
Conclusions
A checkpoint, not yet a product.
Argon remains a working name that surfaced through leaks, routed on Arena under a label that reuses the already-public Gemini 3.8 Flash. Testers report superior speed and aesthetic polish, with concrete examples in 3D rendering compared directly against Claude Opus 5.5. What’s missing is any confirmation from Google, and the October window remains a prediction, not a date. It’s worth watching precisely because it lands in the same month Anthropic and OpenAI are playing their best cards.
To contact us or report updates on this story, use the form at the bottom of the page.
