Riley Brown showed a Gemini 4 Neon session producing a rocket-launch animation in Blender. Then he dropped a clip into Antigravity and asked for timestamped editing notes. Both were convincing reasons to try Neon. Neither was enough to make him move his everyday work out of Claude Code.
That's the useful distinction in this roundup: you can bring a better model into one part of your workflow without replacing the tools that make the rest of it work. Brown liked Neon for code and video, disliked its copywriting habits, and still wanted a better phone-to-workstation experience from Antigravity.
Watch Riley Brown's Update
Source: Riley Brown's November 20, 2026 video and its transcript. The demo results and preferences are Brown's. Listed prices, benchmark scope and product access were checked against public documentation on November 20.
What Neon Costs and What the Benchmarks Measure
Google launched Gemini 4 Neon on November 18 and positions it between Hydrogen for fast execution and Carbon for the hardest long-running work. It lists a one-million-token context window and access through the paid Gemini API, the Gemini app for AI Pro and Ultra users, paid Pro and Ultra accounts in Antigravity, and Gemini for business.
| Model | Input per 1M tokens | Output per 1M tokens |
|---|---|---|
| Gemini 4 Neon | $3 | $15 |
| Claude Opus 5.5 | $4 | $20 |
| Claude Fable 5.2 | $8 | $40 |
| GPT-6.2 Astra | $8 | $40 |
Listed Standard USD API rates before caching, tool fees or higher long-context rates where applicable. Subscription limits are a separate comparison.
That's 25% less per token than Opus 5.5 and less than half the listed rates of Fable 5.2 and Astra. The larger gap Brown discusses comes from a specific test. On Parallel's November 18 search leaderboard, Neon scored 88.1 and cost $233 per 1,000 correct answers, versus $1,220 for Opus, $2,373 for Fable and $235 for Astra. Those are costs with Parallel's search workflow, not general savings on writing or coding. Compare what it costs to finish a real task, including retries and human edits.
What Neon Did in Antigravity
Blender: Brown shows a short rocket liftoff with smoke, camera movement and sound. The visuals were rendered in Blender, according to Brown. That's a useful example of an agent working with a production tool: the model helps construct the scene rather than directly generating the video frames. It is one demo, not a controlled comparison with Fable.
Video review: Brown drops in footage and asks for detailed notes tied to timestamps. Google's Gemini API docs describe video input and timestamped analysis, but also say the default visual sampling can miss quick motion or cuts. That API default does not establish Antigravity's exact settings. Treat notes as an editor's shortlist and verify them against the footage.
Antigravity's launch notes also describe an Orchestrator mode in which Neon plans and Hydrogen handles routine execution such as edits and test runs. That's a different setup from running Neon alone, and it's worth tracking which mode you tested.
Ask for timestamped video feedback
Start with a clip you own or have permission to upload. The prompt above is an adaptation of Brown's demo, not his original wording.
Why the Writing Divided Reviewers
Brown found Neon repeated the same rhetorical move: negate one idea, then immediately replace it with another. It showed up in both an advice response and a short-form script, even when he supplied a style reference. He thought the answers often had substance; the repetitive phrasing was what made them hard to use.
By contrast, Peter Yang narrowly preferred Neon's personal essay to Opus's. Both can be true: a reflective essay and a script in a specific voice are different tests. Try the actual format you publish, hide the model names, and measure how much editing each draft takes. Brown also preferred Opus 5.5 and Fable 5.2 for slides and sites after his early Astra testing; that is a preference from his workflow, not a general finding about GPT-6.2.
Devin Bots: Continuing Work Across Sessions
Cognition describes Devin Bots as persistent coordinators that carry context across engineering sessions, delegate tasks and surface decisions in a review inbox. Brown pairs a bot with iMessage and points out the model picker he saw. Channels and model availability can differ by account.
His standout example is an iOS build on a cloud Mac, followed by a TestFlight link. AWS documents Cognition's macOS environments and their support for building, testing and uploading to TestFlight. A TestFlight build is for beta testing; it isn't an App Store release. For a first run, give the bot a test repository, keep signing credentials scoped and require review before distribution.
Grok Bot: Meetings, Dashboards and Follow-Ups
The official meeting guide covers Google Meet, Zoom and Microsoft Teams on eligible Team, Enterprise, Ultra and Super plans. It describes an approval step, host admission, transcription, and a summary with action items afterward. Brown also highlights team dashboards as a way to see which bots people actually use. The changelog gives the wider release context.
A good first test is a consenting internal call: notify attendees, use an approved workspace and review the notes before sending follow-ups. Brown sees more initiative from Grok Bot and Dots, but a suggestion, a draft and a message sent to a colleague are three different actions. Give each its own permission.
Why Riley Did Not Switch His Whole Stack
At recording, Brown estimated his usage at roughly 75% Anthropic, 20% OpenAI and 5% Grok Bot. Those are his personal estimates, not measured results from a benchmark.
- Claude Code and Projects: Coding and creating decks, spreadsheets or documents.
- OpenAI tools: Voice-led admin and email triage on walks, plus work that requires a browser or desktop.
- Grok Bot: Work that starts on X, including research and sorting incoming opportunities.
Brown wants a first-class phone experience for Antigravity and found the division between Antigravity, AI Studio and the Gemini app unclear. Google already documents browser-based Remote Control, including an installable mobile web app. It can connect to an enabled, reachable host; that's different from the native experience Brown wants.
Try one repeatable job in Neon before moving anything else. Compare the finished result, correction time and how easily you can pick up the session from your phone. If it wins at reviewing footage but loses at drafting in your voice, keep it for footage. A strong model can earn a job without earning your whole stack.