Google's Secret AI CARBON is Apparently a MONSTER

https://bitbiased.ai/ai-automation-services Google employees are reportedly testing an unreleased Gemini 4 checkpoint called Carbon — and one internal reaction says its coding “feels like Opus 5.5.” That sounds significant because Google’s own numbers show Gemini 4 Argon trailing Claude Opus 5.5 by nine points on Terminal-Bench 4.0. So did Google really close one of Argon’s biggest gaps in just a few weeks, or is “feels like” doing most of the work? TestingCatalog reported the Carbon story on October 9, 2026, crediting an original Business Insider report based on internal documents, screenshots and employee messages. According to that reporting, Carbon is one of several internal Gemini 4 checkpoints being tested through Jetski, Google’s internal environment associated with Antigravity. The same leak says Barium-B was the checkpoint that ultimately shipped publicly as Gemini 4 Argon. But the Carbon evidence is extremely limited. There are no benchmark results, code samples, API listings, model cards or public demos. The only performance claim is about coding, and it comes from anonymous internal impressions rather than measured testing. That matters because Argon’s public benchmark profile is unusually uneven. Google reports Argon at 77.9% on DeepSWE, 68.9% on the Vals Index, 65.4% on Finance Agent v2, 91.7% on LVBench and 84.2% on GraphWalks. That last result gives Argon a major reported advantage on very long-context reasoning, matching Google’s positioning around its one-million-token context window. Agentic coding is where the story changes. On Terminal-Bench 4.0, Google reports Argon at 57.4%, GPT-6 Astra at 58.2%, and Claude Opus 5.5 at 66.4%. On FrontierSWE v2, Argon scores 55.0%, behind Opus 5.5 at 62.3% and Astra at 65.5%. That makes the Carbon leak especially interesting. A Gemini 4 checkpoint that genuinely reaches Opus-level coding would address one of the clearest weaknesses visible in Google’s own comparison table. But “feels like Opus 5.5” is not a benchmark. We don’t know the prompts, tasks, tools, agent setup, latency, reliability or evaluation criteria behind the comparison. We also have no Carbon data for reasoning, math, multimodal understanding, video, browser use, long context, safety or cost. And matching Opus 5.5 would not automatically mean Google has matched Anthropic’s highest tier either. The wider comparison is more complicated than a single leaderboard. Argon appears strongest in Google’s reported results on long-context tasks, enterprise knowledge work, video and security. Opus 5.5 leads the Terminal-Bench 4.0 coding comparison. GPT-6 Astra leads several reasoning, math and software-engineering tests. Then there’s pricing. Argon launches at $2 per million input tokens and $10 per million output tokens, but Google says that introductory pricing eventually rises to $4 and $20 — exactly the prices cited for Opus 5.5. And token price alone does not tell us the cost of completing an actual task. Carbon could eventually arrive as an updated Argon checkpoint, become a separate Gemini 4 variant, or never ship at all. Right now, only that last possibility is fully consistent with what is publicly documented: Carbon does not appear in AI Studio, Vertex AI, Google pricing pages or official Gemini documentation. So Carbon should not be treated as Gemini 5, an announced product, or proof that Google has beaten Claude at coding. At this stage, it is an internal Gemini 4 checkpoint that employees reportedly rate highly for coding. The real signal may be what Carbon tells us about Google’s priorities: Argon already has a strong long-context and enterprise story, while agentic coding remains one of its clearest competitive gaps. Until Carbon appears publicly or someone produces an independent Terminal-Bench run, the nine-point difference between Argon and Opus 5.5 remains the only measured coding gap we actually have. :chatgpt-content-reference{index="0"} CHAPTERS 00:00 Google’s Carbon Coding Claim Raises a Big Question 01:36 What TestingCatalog Reported About Carbon 03:07 How Argon, Barium And Carbon Fit Together 04:13 What Google Says Gemini 4 Argon Can Do 05:10 Where Argon Wins In Google’s Own Table 06:40 Where Argon Loses 08:12 What “Feels Like Opus 5.5” Can And Cannot Tell Us 09:22 Argon, Astra And Opus: Three Models, Three Different Strengths 10:30 The Price Catch 11:05 Three Ways Carbon Could Reach You 11:56 Claims About Carbon You Can Safely Ignore 12:29 So What Does Carbon Actually Tell Us? #gemini4 #googleai #claudeopus #gpt6astra #ai