Today for AI
HOT RADAR
llmHEAT 0.0°

AI CLUSTERED EVENT · 10/10/2026

Google Tests Gemini 4 Carbon Model: Coding Capabilities Rival Claude Opus 5.5 on Jetski Platform to Overcome Terminal Bench Bottlenecks

2 reports archived0 independent sourcesupdated 10/10/2026, 16:59:49
Synthesis & Latest Updates

Google is internally testing a new Gemini 4 variant named Carbon, which reportedly rivals Anthropic's Claude Opus 5.5 in coding capabilities. This development directly addresses the performance gap left by Gemini 4 Argon, which scored only 57% on Terminal Bench 4, lagging behind competitors.

Deployed on the internal Jetski platform, Carbon is part of a specialized suite including Argon and Barium. While public benchmarks and release dates remain undisclosed, the fact that Google engineers are currently permitted to use Claude for coding highlights an interim compromise and underscores the urgency of improving proprietary model performance in complex programming tasks.

This move signals a shift towards vertical optimization in the LLM race. If Carbon meets expectations upon launch, it could restore developer confidence in the Google AI ecosystem, mitigate user churn due to weak coding features, and intensify competition with Anthropic in the enterprise code-assistance market.

LATEST/Currently in early internal feedback stage with no public specs or timeline. Google declines comment, while fully opening Claude to engineers, indicating self-developed models have not yet replaced external solutions.

HEAT TRENDHourly heat curve

58 fully observed hours
Now
0.0
Peak
34.117:00
24h change
-100%

The line compares only sources observed throughout; gap hours are interpolated to keep the trend continuous.

TIMELINECoverage timeline

Total 2 reports · Latest first
  2. 宝玉 (@dotey)T1·68 pts
    • Google is internally testing a new model called Carbon, with employees subjectively rating its coding ability as feeling like Claude Opus 5.5.
    • The current model Gemini 4 Argon scored only 57% on Terminal Bench 4, trailing behind Claude Sonnet 5.5 (64%) and Opus 5.5 (60%).
    • Google has allowed all engineers to use Claude for coding, indicating a temporary concession regarding its own models' programming capabilities.