跪拜 Guibai
← All articles
Artificial Intelligence

Gemini 4 in Antigravity: Carbon, Argon, and Hydrogen Benchmarked, Plus Speed Fixes for Slow Code Reads

By oil欧哟 ·
Read original on juejin.cn ↗ Google Translate ↗ Alt translation

Gemini 4's generous quotas and strong communication make it a practical default, but its overconfidence and Antigravity's slow file reads can sink a session. The specific fixes — prompt guardrails, a skip-review plugin, and Swarm dispatching Hydrogen — are what make it usable for real work.

Summary

Google's Gemini 4 family — Argon, Carbon, and Hydrogen — all landed within a month, and a $200 Antigravity plan got a month of heavy use to see how they hold up. Carbon and Argon carry the same API price and nearly identical benchmark numbers; the real difference is a bit more steadiness on coding tasks in Carbon, paid for with noticeably slower responses. Hydrogen is the speed outlier, cheap and fast enough for file exploration, though its engineering judgment is weak.

Against Claude and GPT, the picture is nuanced. Claude Opus 5.5 still wins on complex repository work and default aesthetics; Gemini 4's default UI output is cluttered and ugly unless you constrain it with a design Skill. Gemini wins on communication and Chinese copywriting, and its quota is far more generous than Codex's. But Gemini is dangerously overconfident: it ships plans with absolutes like "100%" and skips verification unless told. The recommended workflow pairs Gemini as the main agent with external Codex for final review.

Speed is the other pain point. Antigravity's harness forces serial file reads through cloud review, turning a complex task into a 10-minute wait. Two fixes: a plugin (agy-fast-io) whitelists commands to skip review, and Swarm can dispatch Hydrogen (Low) as a sub-agent for code exploration. Orchestrator mode exists but is only useful for plan-and-execute, not pure reading — that still runs on the slow planner.

Takeaways
— Carbon and Argon share identical API pricing: $2 input / $10 output per million tokens, doubling to $4/$20 on December 1.
— Hydrogen is $1.50 input / $7.50 output, with a free tier.
— Carbon's benchmark gains over Argon are only a few points on many dimensions; it feels like post-training tweaks, not a generational leap.
— In Antigravity, Carbon reads code for up to 10 minutes on complex tasks because the harness runs serial reads through 1-3 second cloud reviews.
— A global system prompt that bans absolutes and flattery, and forces plain technical language, noticeably changes Gemini's output.
— The agy-fast-io plugin whitelists commands to skip cloud review, cutting file-read time.
— Antigravity's default sub-agent options (inherit/flash_lite/flash/pro) don't include Gemini 4; Swarm can dispatch any model, including Hydrogen with Low thinking.
— Orchestrator mode runs planning on a strong model and hands code changes to Hydrogen; it offers no speed boost for pure code reading because that happens in the planning phase.
Conclusions

Gemini's benchmark scores say little about its overconfidence; a model can ace coding evals yet deliver half-baked code that needs a second model to review.

The real bottleneck in Antigravity isn't token speed but the harness's serial, review-gated file read design; that's a product-level cost that plugins and sub-agent dispatch only partially patch.

Pairing Gemini as the planner with GPT as the reviewer inverts the usual "strong model does the work, small model checks it" pattern; here the expensive Claude/GPT is the safety net because Gemini won't self-verify.

Hydrogen's speed edge is real but its reliability drops in context-rich tasks, so it's best used as a narrow sub-agent for file exploration rather than a general coding assistant.

The aesthetic regression from Gemini 3 to Gemini 4 suggests Google traded visual sophistication for other capabilities; constrained generation becomes a required practice.

Concepts & terms
Antigravity
Google's AI coding tool that provides a subscription plan and hosts Gemini models for code tasks, with features like global prompts, sub-agents, and plugin support.
Harness
The underlying scheduling and execution framework in Antigravity that manages file I/O, cloud review gates, and agent steps, often limiting speed.
Swarm
Antigravity's built-in multi-agent coordination tool that lets you dispatch any model available in the client, including Gemini 4 Hydrogen, as a sub-agent for specific tasks.
Orchestrator mode
A plan-then-execute mode in Antigravity where a strong planning model handles conversation and design, while a cheaper, faster model (Hydrogen) executes code changes.
Skill
A user-defined capability or instruction set in Antigravity that an agent can trigger for specific tasks, such as design or code exploration.
TPS
Tokens per second, the output generation speed of a model; low TPS contributes to slow responses in Antigravity.
Source: juejin.cn ↗ Google Translate ↗ Backup ↗