跪拜 Guibai
← All articles
AIGC · AI · Google

Google's Neon Model: A Comeback Backed by Compute, Price, and Quotas

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

Neon's combination of capability, speed, price, and quotas makes it a viable default for developers who need high-volume AI calls without enterprise budgets. Google's own silicon removes the compute bottleneck that typically forces compromises, and the model's strong showing signals a third major option beyond OpenAI and Anthropic.

Summary

Google released the Neon model last night, and early testing suggests a decisive return to the top tier of AI models. The model isn't necessarily first in every benchmark, but it combines strong overall capability with fast inference, aggressive pricing, and generous usage quotas. That combination, plus Google's own TPU infrastructure, means users are unlikely to get throttled or banned for normal use.

The release follows two months of accelerated model updates after internal restructuring. One developer who has been critical of Gemini's recent quality says the team has been adapting to Antigravity, Google's AI workbench, and the Neon release finally delivers on Google's potential.

The main weakness is Antigravity's user experience, which remains rough. The recommendation is to tolerate it for now and migrate primary workflows to Agy once it upgrades. A practical tip: set up a primary Gmail account through official channels, avoid unofficial top-ups or reverse proxies, and build account history with Google's ecosystem, which becomes more resilient over time.

Takeaways
— Google released the Neon model, which the author rates as overall first among current models when factoring in capability, speed, price, and quotas.
— Neon is not number one in every individual benchmark, but its combined profile makes it a strong practical choice.
— Google's in-house chips give it a compute advantage, so it can offer generous usage limits without banning normal users.
— Antigravity, Google's current AI workbench, has a weak user experience that is a known short-term trade-off.
— The author plans to move all primary workflows to Agy after its next upgrade, and recommends others get an Agy account soon.
— Use a primary Gmail account for Google AI services, avoid unofficial payment or proxy methods, and let account history accumulate.
— Google has not resorted to distilling models from OpenAI or Anthropic, even when Gemini lagged, preferring first-principles development.
— Internal restructuring over the past two months has visibly accelerated Google's model update cadence and relative progress.
Conclusions

Google's willingness to tolerate short-term user experience pain in Antigravity suggests they're prioritizing model quality and infrastructure over product polish, a calculated trade-off for now.

The emphasis on official Gmail accounts and avoiding third-party workarounds points to a real risk: Google's account integrity policies can lock out heavy AI users, so building a trusted account is becoming part of the tooling.

A model that is 'first overall' when pricing and quotas are included shifts the default calculus for developers, who often pick models based on token cost rather than raw benchmarks.

Claiming Google 'never distills' is an opinion, but the absence of known distillation efforts, even when Gemini lagged, is a notable sign of long-term confidence in their own research pipeline.

The mention of RSI (Recursive Self-Improvement) inside Google, even as speculation, reflects a broader industry expectation that frontier labs are moving toward self-improving training loops.

Concepts & terms
Agy
Short for Antigravity? The article uses 'Agy' to refer to Google's Antigravity AI workbench, which is an integrated environment for building with Google AI models. The author plans to migrate to it once it upgrades.
RSI
Recursive Self-Improvement, a speculative scenario where an AI system improves its own abilities iteratively. The author suspects Google may have implemented this internally, though no official confirmation exists.
Source: juejin.cn ↗ Google Translate ↗ Backup ↗