跪拜 Guibai
← All articles
Backend · Programmer · Artificial Intelligence

Gemini 4 Neon Flips the Coding Agent Benchmarks: Faster Than Carbon, Cheaper Than Opus 5.5

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

Agent-heavy coding workflows now have a cheaper, faster option that beats Claude Opus 5.5 on several core benchmarks, and the per-task cost gap is nearly half. Google keeps shipping at a pace that forces teams to re-evaluate their default model choices within weeks, not quarters.

Summary

Google DeepMind rolled out Gemini 4 Neon, the fourth Gemini 4 model in seven weeks, slotting between the heavyweight Carbon and the fast Hydrogen. Google positions it as the high-frequency complex coding and knowledge work option: reasoning close to or above Carbon, but with roughly double Carbon's output speed (about 110 tokens/s) and about 40% less average task time. On the Artificial Analysis Intelligence Index it scores 61 in xhigh mode, ranking third globally and edging out Opus 5.5, GPT-6.2 Sol, and even its own heavier Carbon.

Takeaways
— Gemini 4 Neon is Google's 4th Gemini 4 model in 7 weeks, after Argon, Carbon, and Hydrogen.
— Neon scores 61 on the Artificial Analysis Intelligence Index (xhigh), ranking 3rd overall and above Claude Opus 5.5, GPT-6.2 Sol, and Gemini 4 Carbon.
— Antigravity CLI running Neon (xhigh) tops the AA Coding Agent Index with 59 points, ahead of Claude Code with Fable 5.2 and Opus 5.5.
— CursorBench 4.1: Neon at 65.4% vs Opus 5.5 at 58.1%; SWE-sweep with Antigravity harness: Neon 39.0% vs Opus 5.5 10.1%.
— Neon leads GraphWalks (256K-1M tokens) at 95.2%, 28.4 points above Opus 5.5, and takes first on Zapier AutomationBench and Vending-Bench 2.
— Neon output speed is about 110 tokens/s, nearly twice Carbon, with about 40% less average task time.
— Per-task cost for Antigravity + Neon (xhigh) is $6.86 vs $12.61 for Claude Opus 5.5 (max); default High tier is 32% cheaper than Carbon.
— Neon scores 1771 Elo on AA-Briefcase presentation quality, with a single slide costing $4.03 versus $30.06 for Opus 5.5.
— The new Orchestrator (Neon) lets Neon plan and Hydrogen execute in the same task round, with automatic handback when Hydrogen gets stuck.
— Neon's API pricing is $3 per million input tokens and $15 per million output tokens, with flat 1M context pricing and cache-hit input at $0.15 per million tokens.
— Google demoed Neon in Antigravity with /swarm: building a 3D neon golf game with 8 agents, recreating a Blender physics lab, a WebGL2 ski simulator, restoration of Turing's 1936 study, and running edge detection in Google Sheets with 500,000 formulas.
Conclusions

Google's release cadence is outpacing any single product cycle: 4 models in 7 weeks means the Gemini 4 line is iterating like a SaaS product, not a traditional model series.

The Orchestrator pattern reveals the direction of agent architecture: separate planning and execution models relay control based on task stage, which is a concrete approach to cutting agent failure rates.

Google's decision to bundle Neon into existing AI Pro plans already in use signals a pricing strategy aimed at converting Anthropic users who are fed up with bans and support.

The AA-Coding-Agent top score comes with a catch: it uses Google's own Antigravity harness, so the gap might shrink on neutral harnesses despite the mini-SWE lead.

A $4.03 slide versus a $30.06 slide makes quality-per-dollar the real competitive battleground, not raw benchmark numbers.

Concepts & terms
Orchestrator (Neon)
A Google Antigravity mode where two models work in the same task round: Neon plans and verifies in read-only mode, then Hydrogen executes file changes and tests; control switches back to Neon when Hydrogen fails.
SWE-sweep
A benchmark that evaluates how well an AI agent can resolve software engineering tasks from a repository, measuring end-to-end bug fixes and code changes.
xhigh thinking mode
An extra-high reasoning effort setting available in Gemini models that spends more inference compute to improve problem-solving, at the cost of higher latency and price.
/swarm
An Antigravity command that spawns multiple AI agents to work in parallel on decomposed sub-tasks, coordinating through a Swarm Graph and @mentions.
Steerability
A metric measuring how well an AI agent follows mid-task corrections and direction changes from a user without losing the original goal.
Source: juejin.cn ↗ Google Translate ↗ Backup ↗