Google Gemini 4 Neon Launch: Ranked Third in Intelligence, Coding Task Costs Nearly Half That of Competitors

nashnova research
今天发布阅读约 11 分钟

Google launched Gemini 4 Neon, ranked third globally on an independent intelligence index, with coding-task costs at roughly $6.44 — less than half of Claude Opus 5.5. This means the AI race is shifting from 'who's smartest' to 'who's smart *and* cheap.'

01

What exactly is Neon's positioning?

Neon sits between the flagship Carbon and the lightweight Hydrogen — what Google calls the "sweet spot": roughly 25% cheaper than Carbon, significantly stronger than Hydrogen.
This means → it isn't chasing benchmark crowns in a lab. It targets enterprise-grade daily workloads — complex coding, knowledge work, research tasks.
This is the fourth Gemini 4 model in a single month, following the periodic-table naming scheme: Hydrogen, Argon, Carbon, now Neon. The pace alone has drawn wide attention.
02

How strong is it at coding?

On the Artificial Analysis Coding Agent Index, Neon scored 58.6 — ranked first, roughly matching Claude Code paired with Fable 5.2 and Opus 5.5.
Independent testing in the Antigravity CLI environment yielded 59 points, three above Carbon. Output speed hit roughly 110 tokens per second — nearly double Carbon.
In plain terms = the speed gain doesn't come from cutting corners. Neon actually outputs ~15% more tokens than Carbon; it's faster because the per-token price dropped and throughput doubled.
03

How much cheaper, and where?

One coding task costs Neon roughly $6.86 on average. Claude Opus 5.5 runs $12.61; Claude Fable 5.2 runs $15.47. Neon is less than half of either.
Standard API pricing: $3 per million input tokens, $15 output — about 25% below Carbon's list price.
One caveat: Carbon currently has a promotional rate ($2 input / $10 output) that expires November 30. Neon's full price advantage only kicks in after that.
04

Four thinking tiers — flexibility with a hidden catch?

Neon supports four adjustable thinking tiers: low / medium / high / xhigh. Users trade off speed, reasoning depth, and token spend.
This means → the default tier is high. Without a manual downgrade, real-world costs may be higher than expected.
On the agent side, Neon can act as a dispatcher, coordinating multiple Hydrogen sub-agents in parallel. Google specifically optimized it for long-running tasks — reducing redundant file reads, improving progress feedback, and compressing context. These are well-known pain points in long-horizon agents.
05

How safe and reliable is it?

Google released a Neon cybersecurity edition the same day, covering vulnerability discovery, risk validation, and code remediation — with some test costs roughly half those of large frontier models.
Hallucination rates hit a Gemini-family low on Antigravity internal tests. On indirect prompt-injection tests, the highest thinking tier achieved a 0% attack success rate.
Google itself added a caveat: this result reflects a specific benchmark and does not mean it can withstand all real-world attacks — a level of candor that is unusually rare in the industry.
06

Can benchmark leads translate into real revenue?

Also launched the same day: Stitch Loop — a long-running product agent that connects to enterprise code repos and continuously submits GitHub PRs without auto-merging — plus Gemini App support for low-memory Android Go devices, and a limited rollout of the world model Genie 4 to AI Ultra subscribers.
This reflects a strategy that goes beyond model benchmarks: Google is pushing API + toolchain + device-edge coverage simultaneously.
In plain terms = the core validation question is clear: can coding costs at roughly half the competition's price actually get enterprise clients to switch providers — not just lead on benchmarks?

市场有风险,内容仅供研究参考,不构成投资建议。