Two pinball tables make Greg Isenberg's case for Gemini 4 Neon better than any leaderboard. They combine illustrated playfields, multiball physics and sound generated in the browser. A prospective client can actually play with the result. That is the opportunity in his episode: use Neon to put something in a customer's hands while the idea is still cheap to change.
For a first build, try Neon on High in Antigravity, give it one clear outcome, and make it demonstrate that the result works. A swarm can help when tasks genuinely run in parallel. It won't tell you whether anybody wants the thing you built.
Watch Greg Isenberg's Neon Breakdown
Source: The Startup Ideas Podcast episode and its transcript. Greg says he recorded on November 19, 2026; the YouTube episode was released November 20. Google sponsored the episode. The demos are credited to their creators. Access, pricing and benchmark figures were checked on November 21.
What the Community Demos Show
- Motion design: Chetaslua's Pelican product animation gives a useful reference for product intros, mascots and launch pages. Greg's point is that Neon writes the code that animates the scene; it isn't a native video generator.
- Robotics simulation: The Bugged Dev changed the terrain under a simulated robot dog and showed how it adjusts individual paw placement and body movement. It's a useful example of motion responding to a changing environment, not validation of a physical robot or a design ready to manufacture.
- Pelican on a bicycle: Greg preferred the motion in Neon's version in Bang's side-by-side with GPT-6.2 Astra. That is a judgment about the demonstrated output, not a controlled physics or aesthetics benchmark.
- Pinball: Andrew Voirol shared Nocturne Observatory and a Tokyo-alley table. The builds combine Blender assets, WebGPU rendering, playfield art, multiball and procedural audio. For a pitch, that distinction matters: visuals get attention, but the game has to respond to controls.
Where to Run Neon
| Environment | Use it for | What to check |
|---|---|---|
| Gemini app | Quick visual concepts, drafts and work grounded in Google apps you have authorized. | Google's launch notes list Neon for Google AI Pro and Ultra users. |
| AI Studio / Gemini API | Multimodal inputs such as long PDFs, audio or video; instrumented prompt tests. | The API docs list 1,048,576 tokens shared by input and output. The standard paid rates are $3 input and $15 output per million tokens; there is no free API tier for Neon. |
| Antigravity | Multi-file builds and browser checks where the agent needs to edit, run and inspect the project. | The Antigravity model docs list up to 512K context for Neon; supported plans and settings vary. |
Google reported 12.8% in mini-SWE and 39.0% in Antigravity on SWE-sweep. Both results use Neon, but the software around it changed: how the model receives tools, works on the repository and checks its work. That is a 26.2 percentage-point increase on this coding benchmark, not a promise that every project will improve threefold. Google's model card describes the evaluation.
Five Ways to Get Better Results
- Start on High. In Artificial Analysis's test, Neon High scored 60 at $3.74 per task; Extra High scored 61 at $6.44, about 72% more for one index point. Google separately reports a 32% cost-per-task advantage over Carbon at Neon's default High setting. Neither comparison is a flat token discount or a forecast of your bill.
- Use Antigravity for a build, not just an answer. Give the agent a project it can run locally and an acceptance test that describes what the user should be able to do.
- Separate Orchestrator from Swarm. In Google's Neon Orchestrator pairing, Neon plans and Hydrogen handles routine execution.
/swarmcoordinates multiple agents when tasks can run in parallel. Greg suggests research, design, frontend and backend roles. Start smaller if they would all touch the same files, and keep deployment behind approval. - Ask for proof the result works. Have the agent load the interface, use the controls, test failure cases and fix what breaks. Ask for a screenshot or short recording and a list of checks it actually ran.
- Put stable context first. Lead with the brief, reference material and facts that shouldn't change, then add the request. Google's implicit caching guide says a common prefix can improve the chance of a cache hit. It's neither guaranteed savings nor a substitute for checking sources.
Five Business Ideas Worth Testing
| Idea from the episode | First pilot and what to measure |
|---|---|
| Playable lead magnet | Build a small calculator, game or configurator that gives visitors a useful result before asking for an email. Compare qualified leads with a simpler page. |
| Custom prospect demo | Make a private, clearly labeled mockup for one business, using authorized assets or placeholders. Show one useful improvement and track whether it earns a follow-up conversation. |
| Margin and pricing simulator | Use redacted or synthetic invoice data to build a model. Reconcile totals against a trusted sheet and have the finance owner check the assumptions before relying on any result. |
| Back-office automation | Start with read-only lead routing suggestions or a reconciliation queue. Measure minutes saved and errors caught before granting any ability to change records. |
| Micro-product or directory | Pick one audience and let separate tasks cover research, the prototype and review. Check whether five real users return before investing in polish or a larger build. |
Greg floats low monthly fees for small-business redesigns and contrasts them with far larger agency projects. Those are examples, not verified market rates or revenue forecasts. Quote the scope you can actually support, including hosting, accessibility, analytics and maintenance where relevant.
Build and verify a playable lead magnet
This prompt adapts Greg's idea. It is not a transcript quote.
Four Caveats Before You Switch
- Brevity: Greg found some Neon answers much shorter than Fable's. Ask for the evidence and depth you need rather than simply asking for a longer answer.
- Writing habits: He noticed repeated em dashes and stock contrast phrases. Give the model one example of your voice and tell it which patterns to avoid.
- Naming and access: Google positions Hydrogen for fast execution, Neon for demanding everyday work and Carbon for the hardest projects. The same model name does not mean the same context limit or billing in every product.
- Antigravity's rough edges: Greg found the built-in browser useful but imperfect. Recheck the main interaction in your own browser, and compare the time spent reviewing and fixing it with your usual setup.
Run one small job you've done before. Keep the brief and acceptance test the same, then compare time to an output you'd actually ship, including retries and cleanup. That comparison will tell you more than a spectacular one-off demo.