A generated title card reading 'Gemini 4 Neon's Stickman Game' with the subtitle 'the swarm claim, not the demo', beside two cards: on the left 'The artefact' listing a 62.5-second video at 756 x 403 playing a browser game with no prompt and no settings recorded, on the right 'The record' listing three dated Google posts — 10 Nov /swarm ships, 12 Nov Orchestrator ships, 18 Nov Neon ships — over a footer reading 'One posted clip; three dated Google posts; one untested training claim.'
Guides & Insights

Gemini 4 Neon's Stickman Game: The Swarm Claim, Not the Demo

Author

jinhao.song

Date Published

Latest models · 20View all models →
Benchmarks: Artificial Analysis · updated daily
Back to all posts

At 20:41 UTC on 19 November 2026, an account posting as @chetaslua put up a 62.5-second video of a two-player stickman fighting game and four sentences of text. The game, it said, was “Made by swarms of Gem​ini 4 Neon, pure JS code.” The model, it said, “has distinct taste and is different from O​penAI and A​nthropic in a good way.” And then the load-bearing line: “The benefits of not doing distillation and taking time with RSI are showing now.” Gem​ini 4 Neon had been generally available for one day at that point, having gone out on 18 November, with Gemini 4 Carbon and Gem​ini 4 Hydrogen already shipping ahead of it. So this is not a leak and not a preview. It is a claim about how the model was built and how it was used, published by someone who is not G​oogle, one day after launch, with a video attached and nothing else.

Most of what has been written about Neon so far has been about its numbers, and the numbers are genuinely interesting: 61 on the Artificial Analysis Intelligence Index at its Extra High setting, first place on the Coding Agent Index at 58.6, and a default effort level that undercuts Carbon on cost per task. This piece is about the other half of the post — the half that says the model is the product of swarms and of a training philosophy G​oogle has not described anywhere. Those are separable claims. One of them turns out to be almost entirely documented, by G​oogle, with dates. The other has no public receipt at all, and the difference between them is the useful part.

What 'made by swarms' actually names

There is a temptation to read “swarms” as a disclosure about training: a model built by fleets of copies of itself, recursively. Nothing in the post supports that, and everything in G​oogle’s own publishing points somewhere else. There is exactly one thing in the Gem​ini 4 record called a swarm, and it is a product feature, not a training method.

On 10 November 2026, the Antigravity team published a post titled “Gemini 4 Carbon and /swarm.” The relevant sentence is a description of the harness, not the model: /swarm “passes your prompt to a family of agents: a coordinator that spins up specialized parallel workers in isolated git worktrees so edits never collide, and adversarial verifiers dedicated to running tests and challenging diffs.” Every agent collaborates on “a shared, append-only message board, parking at zero token cost when idle and activating the instant an agent needs input.” Agents can be pinned to different models, so a coordinator can reason on one model while workers write on another.

That is a swarm in the ordinary engineering sense: parallel processes with a shared queue and a verification step. It says nothing about how Gem​ini 4 Neon was trained. Whether the poster meant that, or meant something looser, is not recoverable from four sentences — but if you are trying to place the claim against the record, the record’s swarm is a dated Antigravity feature and nothing more.

The two-player detail G​oogle published

Here is where the post stops being vague in an interesting way. The Antigravity post announcing Gem​ini 4 Neon, published 18 November, contains a captioned demo video whose caption reads: “Using Gem​ini 4 Neon with /swarm, Antigravity scaffolds a multi-level 2-player game that runs in the built-in browser while the agent plays as the second player.”

Read that next to the tweet. A multi-level two-player game, scaffolded by a swarm, with the agent as the second player, in a browser, from the vendor’s own documentation, published a day before the video. The tweet does not say it came out of Antigravity — it names neither the app, the API nor AI Studio, and gives no prompt, no effort level, no token count and no source file. But the one publicly demonstrated swarm path to a two-player browser game is Antigravity’s, and it is documented as working with this exact model.

That does not verify the clip. A vendor’s captioned demo and a third-party video are different kinds of evidence, and neither is a specification. What it does do is make the interesting part of the post specific: “made by swarms” is not a claim about training at all, if it means what G​oogle’s own material means by the word. It is a claim about a harness that has been shipping for nine days.

Three dates that document the swarm

The swarm story has a timeline, which is more than most launch-week claims get. All three entries are G​oogle’s own, from the Antigravity blog:

• 10 November 2026 — Gemini 4 Carbon ships with /swarm in preview. Coordinator, parallel workers in isolated worktrees, adversarial verifiers, shared message board, model pinning per role.

• 12 November 2026 — Gem​ini 4 Hydrogen ships with Orchestrator, a dropdown option that hands work back and forth between Hydrogen and a stronger model inside a single trajectory. The post claims near-Carbon task success on Terminal-Bench 4.1 while consuming 24% less quota and finishing 19% faster, and notes that “tens of thousands of swarms” had been launched in the two days since /swarm went live.

• 18 November 2026 — Gem​ini 4 Neon ships, and Orchestrator gains a Neon-as-planner, Hydrogen-as-executor pairing. The same post claims Neon at its default High thinking level runs 32% cheaper per task than Carbon, and puts it at 61 on the Artificial Analysis Intelligence Index at Extra High.

A generated single-column scoreboard titled 'Gemini 4 Neon — the scoreboard' with six rows: swarm harness /swarm, preview 10 Nov 2026; Orchestrator role planner, paired with Hydrogen; AA Intelligence Index 61 at xhigh; default effort high, 32% cheaper per task; cost per index task $6.44; training disclosure none published. A footer reads 'Per Google Antigravity posts and Artificial Analysis; RSI and distillation claim unaudited.'

That is the whole of the public account of Gem​ini 4 swarms, and it is a product account. Note what is absent from all three posts: the words “distillation,” “self-improvement,” “reinforcement learning” and “post-training” do not appear. The Orchestrator paragraph about routing is the closest any of them comes to describing model behaviour as a system rather than a model, and it is explicitly about inference-time model selection, not training.

Screenshot of the Artificial Analysis page for Gemini 4 Neon, headed 'Intelligence, Performance & Price Analysis', showing an Intelligence Index of 61 ranked 6 of 95, $3.00 input and $15.00 output per million tokens, $6.44 cost per Intelligence Index task, 21.5 minutes per task against a 12.4-minute median, 283 million output tokens during the evaluation against a 65.4-million median, a verbosity rank of 92 of 94, and a 3.482% output-to-input token ratio.

The RSI claim arrives without a receipt

The second half of the post is the part with no supporting document behind it, and it is worth being precise about why, because the claim is not obviously wrong — it is unverifiable in a specific way.

Start with the acronym. In this literature “RSI” is used for at least two different things: recursive self-improvement, meaning a model participating in the improvement of its successor, and rejection-sampling iteration, meaning repeated rounds of sampling, filtering and fine-tuning on the survivors. “Taking time with RSI” reads more like the first. Either reading is an assertion about the training pipeline, and G​oogle has published nothing that would let you choose between them, let alone confirm one.

Check the pages that would carry it. DeepMind’s Gemini model page lists Gem​ini 4 Neon under “November 2026” with no day, and discusses capability, price and availability; it does not mention distillation, self-improvement, reinforcement learning or post-training anywhere. The Antigravity post is about harness integration. The model card is the place a training-process claim would normally surface, and the Neon card does not carry one: seven of its slots print the same clause, “Gem​ini 4 Neon is adapted from Gemini 4 Carbon,” and point at the Carbon card, with no parameter count and no occurrence of “distill,” “prune” or “quantize.”

That last point puts the two halves of the tweet in tension with each other. “Adapted from Gemini 4 Carbon” is a statement of derivation. A model adapted from another model is not, on the face of it, a model built without distillation — unless “adapted” is doing work none of us can see. The post asserts the absence of a technique that the only official document about the model’s provenance declines to rule in or out. It could all be true. It is not checkable, and the person making the claim has not shown the working.

The falsifiable part of the same sentence is easier to test, and it is where the numbers do some work. If Neon really is the product of a patient, non-distilled training process, you would expect the payoff in capability per token rather than in speed. What the independent tracker shows is a model that is capable and unusually expensive to run: 61 on the index at Extra High, but $6.44 per index task, 21.5 minutes per task against a 12.4-minute median, and 283 million output tokens generated during the index run against a 65.4-million median — a verbosity rank near the bottom of the field. That is a profile consistent with a model that thinks for a long time. It is also consistent with plenty of other things. It does not settle the training question either way, and no published figure does.

What the clip does and does not establish

Strip the philosophy out and the artefact is thinner than the argument around it. The video is 756 by 403, runs 62.5 seconds, and shows a two-player stickman fighter running in a browser. There is no prompt in the post, no model version string, no thinking level, no token accounting, and no statement of whether one agent or forty produced the code. “Pure JS code” is a claim about the output language; it is not a claim about how the code was produced, and it is not a measurement of quality.

What the clip does establish is bounded but real: a browser-playable two-player game in plain JavaScript can be produced on this model through a swarm harness, and someone did it within a day of general availability. That is a capability demonstration, and the vendor’s own post corroborates the shape of it. What it does not establish is any statement about training, because a demo cannot carry one — the causal chain from a training process to a stickman game runs through too many decisions to be visible in a video, and nobody involved has published the settings that would let you re-run it.

This is where the honest reading of the whole post lands. The swarm half is documented, dated and reproducible in principle. The training half is a claim by an observer about work they did not do, published without the one thing that would make it checkable, which is a source. Treat the first as a description of a shipped feature and the second as a hypothesis in circulation, and the post becomes useful instead of merely interesting.

A version of the test you can check

If the swarm claim is a harness claim, it can be tested on your own traffic, which the training claim cannot. The pieces you need are a coordinator, workers, and a verifier — and the cheapest version is to pin each role to a model you can already call.

• Record the brief — the prompt, the effort level, and the model pinned to each role. Without these three, a re-run is a different experiment.

• Measure the loop, not the artefact — total tokens, wall-clock time, and how many verifier rounds the diff survived before it passed.

• Keep the verifier adversarial — the Antigravity design pairs a builder with something dedicated to breaking its output, and that second role is doing most of the quality work.

One practical note on where this can run today: Gem​ini 4 Neon is not in the OrcaRouter catalogue, which currently carries 217 model identifiers and, among the Gem​ini 4 family, only google/gemini-4-carbon and google/gemini-4-argon, both listed at $2.00 in and $10.00 out per million tokens with cache reads at $0.10. So a coordinator-worker split on our side means splitting across those two rather than pinning Neon, and the routing DSL is the place to express it — compose a planner and an executor into a single call, and the failover path keeps a dead worker from taking the run with it. For comparison, the same Neon traffic at its published rate is $3.00 and $15.00 per million, with batch and flex at half, which is the number to hold against any harness you build.

Screenshot of the OrcaRouter model page for google/gemini-4-carbon, showing the model header labelled by Google dated 2026-11-10, a 1,048,576-token context window accepting text, image, audio, video and documents and returning text, both the /v1/chat/completions and v1beta generateContent endpoints, input priced at $2.00 and output at $10.00 per million tokens, a 10.60 s performance figure, zero-data-retention availability, and an OpenAI-compatible Python sample calling google/gemini-4-carbon through https://api.orcarouter.ai/v1.

The thing worth watching is not the next clip. It is whether anyone — G​oogle, or the person who posted this — publishes a tool-call log or a token count from a swarm run on Gem​ini 4 Neon at a stated effort level. That would turn a demo into a measurement, and it would let the swarm claim be checked the way a shipped feature should be. The RSI claim needs something different: a training disclosure that does not currently exist on any G​oogle page, and until one does, the most accurate thing to say about it is that it is unconfirmed rather than that it is false.