Gemini 4 Neon shrank the TS Rust binary from 47.44 to 7.97 MB in under 10 hours
Google released the mid-tier model Gemini 4 Neon, and blogger Theo put it to the test on TS Rust, a TypeScript type checker written in Rust. In 10 hours, Neon shrank the finished program file from 47.44 to 7.97 MB (5.95 times smaller) and made it 1.5 times faster on average. GPT-6.2 Astra worked on the task for over a week: the file came out 46% smaller, but the program ran 56% slower. This is one author's video, so you should test the model on your own code before trusting it.
AI-processed from Habr AI; edited by Hamidun News
Imagine a task developers usually put off: there is a large project, the finished program weighs 47 megabytes and runs slower than its competitors, and you need to cut its size and speed it up. On November 18, Google quietly released the mid-tier model Gemini 4 Neon, and in a video test by blogger Theo, it handled this task on the TS Rust project (a TypeScript type checker written in Rust) in under 10 hours. The finished program file, the binary, shrank from 47.
44 to 7.97 MB, which is 5.95 times smaller, and its speed rose 1.
5 times on average. For comparison, GPT-6.2 Astra from OpenAI worked on the same task for over a week and made the program 56% slower.
The author of the Habr review concludes from this that Anthropic's Claude Opus 5.5 now has an equal rival. An important caveat: this is the author's personal experience and one blogger's video, not an independent benchmark.
But it contains enough concrete detail to analyze.
What the TS Rust test showed
The most striking episode comes from a 32-minute video by developer Theo. He had spent half a year criticizing Google's models, but here he gave several models the same task: shrink the binary and speed it up to catch up with the competitors bun and tsrs.
According to the author's account, the results look like this:
- GPT-6.2 Astra worked for more than seven days and then spent another four days trying to fix the outcome. The binary became 46% smaller, but speed dropped by 56%.
- Claude Opus 5.5 together with Fable finished in a few days and gave about +20% speed, but without significant compression.
- Gemini 4 Neon, in swarm mode (several agents working in parallel), finished in under 10 hours. Speed rose 1.5 times on average, 2.3 times on the Sentry project, and up to 20 times on a warm start, when the program is launched again and data is already in memory.
Two more episodes from the video show agent behavior. When a broken MCP adapter swallowed the output and nothing appeared in the chat, Neon went into the logs by itself, found the bug, rendered the charts locally, showed screenshots, and opened a PR with the fix. And when searching for the reason why a git worktree took 6.3 seconds to create, it found the real source: the /work partition had been formatted for a regular HDD, and this setting accounted for 3.8 of the 6.3 seconds.
Prices, limits, and cache
Neon is positioned as a mid-tier model built on Hydrogen. API price: $3 for input and $15 for output. Hydrogen costs $1.50 and $7.50, and the top-tier Carbon has a discount through November 30 ($2 and $10); from December it returns to $4 and $20. So in November the "mid-tier" model is more expensive than the "top-tier" one, and from December it will be 25% cheaper. Names in the style of the periodic table add to the confusion.
The subscription story is separate. On paper, the weekly budget of Google AI Ultra 20x is estimated at $2000, compared with $3000 for Claude Ultra and $1000 for ChatGPT Pro. But in the first days, prompt caching worked poorly: after a pause of less than a minute, cache hits dropped to 9% or to zero. Every request reprocessed the entire context at full price, and limits burned about 2.5 times faster.
After complaints on X, the head of Antigravity confirmed the fix was rolling out to all clusters and promised to restore the spent limits. According to the author, the cache now lives for about three minutes, and real limits rose roughly threefold. Antigravity also gained Orchestrator: Neon plans and handles the complex parts, while routine edits and tests go to the cheap Hydrogen. If you step away for coffee for five minutes, the context still has to warm up again, so the limit still needs to be accounted for.
Where Neon is weaker and what to do
The author honestly lists the drawbacks:
- the visual frontend is noticeably simpler than Opus 5.5's;
- in 3D, the model trails both Opus and OpenAI's heavy models;
- "Geminisms" occur, such as extra emojis and fragments of internal reasoning;
- image generation and parsing raw web transcripts are better left to ChatGPT.
In practice, the takeaway is this. If you build agentic workflows around code, review, and analysis of infrastructure problems, Neon is worth adding to your comparison rather than relying on a single provider. It is also insurance against sudden blocks, which the author complains about in relation to Anthropic. But test the model on your own repositories: one video and one review do not replace your own measurements. Also check quota consumption and cache behavior in your own working mode. And note that Antigravity's system prompt, by default, makes the model clutter its answers with long file paths. This is fixed with a couple of lines in user rules.
If the author's measurements hold up in independent tests, Google's mid-tier model is already handling tasks that used to require expensive flagships. In that case, the cost and speed of agentic development would change noticeably.
- Nov 21, 2026A swimming team captain handed ChatGPT three scheduled tasks and cut her routine work by a multiple
- Nov 21, 2026MIRA Neural Network Renders Rocket League at 20 FPS, but Ping Hits 460 ms
- Nov 21, 2026Google built a closed copy of the internet with fake banks to train Chrome's AI agents
- Nov 21, 2026MLPerf Training Adds Nvidia Rubin and TPU Ironwood in First Appearance: 153 Systems
Key Questions Answered
What the TS Rust test showed?
The most striking episode comes from a 32-minute video by developer Theo. He had spent half a year criticizing Google's models, but here he gave several models the same task: shrink the binary and speed it up to catch up with the competitors bun and tsrs. According to the author's account, the results look like this: GPT-6.2 Astra worked for more than seven days and then spent another four days trying to fix the outcome. The binary became 46% smaller, but speed dropped by 56%.
What are the key details on Prices, limits, and cache?
Neon is positioned as a mid-tier model built on Hydrogen. API price: $3 for input and $15 for output. Hydrogen costs $1.50 and $7.50, and the top-tier Carbon has a discount through November 30 ($2 and $10); from December it returns to $4 and $20. So in November the "mid-tier" model is more expensive than the "top-tier" one, and from December it will be 25% cheaper. Names in the style of the periodic table add to the confusion. The subscription story is separate.
Where Neon is weaker and what to do?
The author honestly lists the drawbacks: the visual frontend is noticeably simpler than Opus 5.5's; in 3D, the model trails both Opus and OpenAI's heavy models; "Geminisms" occur, such as extra emojis and fragments of internal reasoning; image generation and parsing raw web transcripts are better left to ChatGPT. In practice, the takeaway is this.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.