Google's Gemini 4 week gave Wes Roth several different things to test: a frontier model, a faster sibling, a closed-choice decision API, a formal-verification research system and a new multi-agent mode in Antigravity. His 30-minute walkthrough is most useful when those are kept separate. It includes a striking video edit and a Factorio agent, but also a replay bug, a heavy quota bill and a temporary account lockout.
Carbon is Google's deeper, long-running model; Hydrogen is the faster everyday model; Decision 1 scores bounded choices; AlphaProof Code is a human-guided verification framework; and Antigravity's /swarm coordinates parallel agents. Wes's demos show promising workflows, not a controlled proof that one model wins every task.
Watch Wes Roth's Full Test
Five Announcements, Five Different Jobs
| Release | What it is | Where to read more |
|---|---|---|
| Gemini 4 Carbon | Google's highest-effort Gemini 4 model for long-horizon coding, research and complex knowledge work | Launch post · model card |
| Gemini 4 Hydrogen | Faster model adapted from Carbon for everyday coding and agent execution | Launch post · model card |
| Gemini Decision 1 | API model for yes/no, choice and rubric scoring rather than long generated responses | Google's API introduction |
| AlphaProof Code | Research framework combining human-approved specifications, agent-generated proofs and Lean checking | DeepMind report · paper and artifacts |
| Antigravity /swarm | Preview mode for parallel workers, verifiers, isolated worktrees and a shared message board | Launch post · documentation |
Google says Carbon is for its hardest problems and Argon remains the better starting point for most engineering and knowledge work. Carbon's introductory API price is $2 per million input tokens and $10 per million output tokens through 30 November 2026, after which Google lists $4/$20. Hydrogen is listed at $1.50/$7.50 across its 1-million-token context window. These are API list prices, not the cost of Wes's subscription-based runs; verify live terms before budgeting.
The Video Edit: Impressive, With an Important Correction
Wes shows a prior video that he says Carbon edited end to end: cutting silences and retakes, organizing chapters and adding coded graphics, charts and a subscribe animation. His example matters because useful editing requires more than a transcript. The model appears to match overlays to visual and audio context, including music-video footage and the subject of a spoken segment.
The opening description can be read as “Carbon edited this entire 30-minute video,” but the transcript is more specific. At 06:26, Wes says Gemini 4 Argon edited the current upload, while Carbon handled the earlier example. Neither statement includes a shared raw project, edit decision list, timing record or repeat runs. Treat the result as a creator's demonstration. For a production trial, keep the source footage and project files, then review every cut for meaning, visual accuracy, music rights, caption errors and final audio.
Factorio: A Harness, Not Just a Clever Screenshot
Wes asks Carbon to create a way for language-model agents to play Factorio. He says it reused existing mods, built a harness, ran agents through mining and research, and got them to train tracks and trains. The first replay did not work inside the game; a second pass repaired it. Later in the video, an agent is nearly ten hours into a harder rocket-launch run, but has not launched the rocket by the recorded ending. The test therefore supports the narrower claim: Carbon helped build and operate a game-playing harness and advanced a run substantially.
That is still a good evaluation idea. The visible setup mixes model planning, tool design, a game interface, mods and repeated runs. If comparing models, hold those conditions constant, record success criteria up front and count restarts, time, tokens and human interventions. Wes's MaxBench arena is explicitly early: he says it needs more games before its rankings are reliable. Google's published benchmark placements are also Google-reported claims, not substitutes for testing your own workload.
AlphaProof Code: What Was Formally Verified?
Wes treats the Signal result as the week's largest technical news. Google DeepMind reports that researchers used agents powered by Carbon to produce machine-checked Lean proofs for specified security properties of compiled libsignal code. The team approved the trusted specifications and reviewed escalated issues. The published paper and proof artifacts explain the AArch64 setup, assumptions and what is outside its model.
This is more substantial than an AI merely saying a protocol is secure, because Lean checks the proposed proofs against the specified statements. It is not a proof that all of Signal, every deployment environment or every possible security property is flawless. The result depends on the chosen specification, cryptographic assumptions and modeled execution conditions.
/swarm Worked; Unbounded Parallelism Did Not
Google's /swarm documentation describes a coordinator, parallel worker agents, a verifier, isolated Git worktrees and a shared board. Wes likes the visible board and graph because he can see agents coordinating and inspect the run. Google labels the feature a quota-heavy preview; it is available on paid Google AI Pro and Ultra plans, while Carbon access is more limited. The separate Orchestrator option switches between models within one session. It is not another name for /swarm.
In another experiment, Wes says a Claude agent launched roughly 50–100 background Antigravity sessions in quick succession while setting up benchmark tests. His account then stopped accepting requests; he appealed and says access was restored. We cannot independently confirm the precise trigger from the video alone. The operational lesson is clear without bypassing safeguards: cap concurrency, watch quota and spend, keep approval gates on and check the provider's usage policy. Do not treat a tool's ability to run unattended as permission to flood a service.
What to Try First
For a bounded classification or routing job, try Decision 1 and verify its confidence against labeled examples. For fast, repeatable coding, start with Hydrogen. For an ambiguous, multi-hour build, compare Carbon with your existing model on the same brief. Reach for /swarm only when the work really divides into parallel files or tests and you can monitor quota. Use the model cards to check inputs, limitations and evaluation context, then judge by the quality of the finished artifact, not the number of agents on screen.
Video Chapters
| Time | Topic |
|---|---|
| 00:00 | Gemini 4 Carbon |
| 01:00 | AlphaProof Code and benchmarks |
| 02:34 | Decision 1 and model names |
| 04:02 | Gemini 4 Hydrogen |
| 05:08 | AI editing the video |
| 08:05 | Factorio agent test |
| 11:30 | Vocal-tract music experiment |
| 14:59 | Agent blitz and MaxBench |
| 20:54 | Account lockout |
| 27:11 | Antigravity swarm verdict |
Original Sources and Caveats
Based on Wes Roth's video and the supplied transcript, checked against Google and DeepMind publications on 13 November 2026. Video-editing, Factorio, early MaxBench and account-lockout results are Wes's reports. Google benchmarks and product-performance numbers are vendor-reported unless otherwise stated. Availability and prices can change.