谷歌内测Gemini 4 Carbon,编程体验比肩Opus 5.5
据Business Insider报道,谷歌员工正在内部编程平台Jetski上测试代号Carbon的Gemini 4新版本。一名员工称其编程体验“像Opus 5.5”,但强调还需更多测试;谷歌内部已测试Argon、Barium和Carbon等多个版本,尚不确定Carbon会作为Argon的更新还是独立型号推出,谷歌拒绝置评。与此同时,已宣布的Argon尚未广泛开放,Antigravity中出现了相关模型引用和订阅提示。
若员工的主观评价成立,Carbon可能缩小谷歌在编程智能体任务上与Anthropic的差距,但该说法尚未经独立验证,谷歌也未确认。
Google is preparing Antigravity for Gemini 4 Argon, with new labels appearing around usage limits, subscriptions, and prompts to upgrade to Google AI Ultra. These additions suggest Ultra subscribers could be among the first users to access Argon in the coding app. Google’s September 30 announcement already names Ultra subscribers and paid API customers as the starting point for its broader rollout, after giving access to trusted cyber defenders.
UTM links in Antigravity already point to destinations mentioning "Argon" related params: utm_source=antigravity&utm_campaign=argon_model_selector
Google says Argon is already powering internal workflows and reports a 77.9% score on DeepSWE v1.1, which evaluates longer software engineering tasks. This places Antigravity in Google’s push toward agents that can handle complex coding work over extended periods. The new subscription labels point to rollout preparations, although they do not establish an Antigravity release date.
0:00
/0:08
We also spotted an agent prototype internally called Concierge, with support for real-time voice. It appears to be an early implementation, so its presence does not establish that Google plans to ship it. If released, Concierge could let developers direct agent work through spoken requests.
#Google#Gemini 4 Carbon#Gemini 4 Argon#Antigravity#Jetski#Claude Opus 5.5#Google DeepMind#Barium