Huge Gemini 4 Carbon Leaks, Claude Fable 6, GPT-7 Bel, Google Cracked RSI?! & More! AI News

The headline is still Gemini 4 Carbon

The working headline matches the subject: Huge Gemini 4 Carbon Leaks, Claude Fable 6, GPT-7 Bel, Google Cracked RSI?! & More! AI News. The strongest fresh evidence in the story is Business Insider’s October 9 report that Google employees are testing an internal Gemini 4 model version named Carbon, even as the company prepares the public rollout of Gemini 4 Argon [1].

According to that report, Carbon appeared inside Google’s internal coding platform, Jetski, and may be a later checkpoint or update related to Argon rather than a separate public product [1]. The most repeated claim is that one Google employee said Carbon “feels like Opus 5.5” for coding, while also warning that more testing would be needed before making a firm comparison [1]. That caveat matters: there is no public benchmark, no model card and no independent evaluation showing Carbon actually matches Anthropic’s top coding model.

The Carbon leak also adds context to Google’s internal naming pattern. Business Insider reported that Google has tested Gemini 4 builds internally under names including Argon, Barium and Carbon, and that an internal “Barium-B” build was selected to become the public model known as Argon [1]. A Reddit post circulating the same leak framed Carbon as a stronger internal checkpoint and said leaked documents and screenshots indicated deployment to internal software-development platforms for employee testing [2]. Reddit is not confirmation, but it shows how quickly the story moved from trade reporting into developer-community speculation.

Argon may not be the end state

The important interpretation is not simply “Google has another codename.” It is that Google appears to be iterating several Gemini 4 checkpoints in parallel while the public still has not received the full Argon release. HelloBro’s own October 10 roundup described the Carbon claims as unverified, but noted that employees reportedly compare the model’s coding behavior with Claude Opus 5.5 and that earlier codenames such as Beryllium or Barium point to parallel development [3].

That is strategically significant because coding agents have become the most visible frontier battleground. If Carbon is only an internal checkpoint, Google can still use it to improve Argon before launch, harden safety layers, test long-horizon coding workflows and compare real employee usage against public benchmark claims. If Carbon becomes a separate Gemini 4 tier, the story becomes more aggressive: Google would be signaling that Argon is not the ceiling.

The restraint here is essential. Google declined to comment to Business Insider [1]. No public Carbon API exists. There is no official Carbon announcement. For now, the sober reading is: a credible report says Google is testing a stronger Gemini 4 coding checkpoint internally; performance claims remain anecdotal.

Did Google “crack RSI”?

The most explosive phrase in the title is also the least verified: recursive self-improvement, or RSI. In this news cycle, RSI does not mean a confirmed runaway intelligence event. It refers to claims that labs may be using current AI systems to speed up model research, training, evaluation, debugging, data work or architecture iteration.

HelloBro’s roundup says separate insider claims suggest Google may be close to, or already using, forms of recursive self-improvement, with speculation tied to the pace of Gemini Flash progress and Gemini 4 checkpoint iteration [3]. The Japanese ASI translation of the same WorldofAI-based story similarly stresses that the RSI angle is unverified and that rapid model improvement alone does not prove true recursive self-improvement [4].

That distinction is the whole story. A lab using AI to write tests, repair code, evaluate models, search research space or automate data pipelines is plausible and already consistent with the direction of the industry. A lab achieving self-sustaining recursive improvement is a much stronger claim. The current evidence supports the first interpretation far more than the second.

Google’s official enterprise news adds fuel, but not proof. On October 8, Google Cloud introduced the Gemini agent as a “single, universal agent for work” able to answer questions, handle knowledge work, create media, and write and run code from one interface and API [5]. That kind of tool could help enterprises and internal teams automate complex workflows. It does not show that Google has cracked RSI.

Anthropic: Fable 6 rumors meet real workflow changes

The Anthropic portion of the story is also split between rumor and verified product movement. The rumor is that Anthropic may have started a new pre-training run beyond the expected Claude Fable 5.5, possibly aimed at a larger Fable 6 or Opus 6-class model [3]. As of this writing, that remains unconfirmed.

The verified side is more concrete: Anthropic’s Managed Agents are moving toward broader multi-agent orchestration. Public summaries of Anthropic’s October 9 platform update describe dynamic workflows in Claude Managed Agents, where a lead agent can write and run a workflow that coordinates many agents in phases and combines their results server-side [6]. Chasing Next’s summary says Anthropic’s own test on a 116,000-line codebase with 70 planted bugs saw a dynamic workflow find 66 bugs, compared with much lower single-agent runs [6].

That is highly relevant to the Fable 6 rumor even if it does not prove it. Frontier labs are no longer competing only on raw model intelligence; they are also competing on orchestration. If a model can break a task into dozens or hundreds of subtasks, manage specialized agents, and recombine results, it can appear dramatically more capable in real software work. That is why Claude’s workflow layer matters in the same conversation as Gemini Carbon.

OpenAI: GPT-7 Bel remains speculative, Codex changes are real

The OpenAI part of the title centers on GPT-7 Bel or Bell, a rumored next-generation model that some communities expect could arrive by December. Current public evidence does not confirm that name, launch window or capability level. A Reddit thread framed the claim as coming from a supposed Google employee and included debate over whether Bel would beat upcoming Anthropic and Google releases [7]. That is not a reliable announcement.

What is real is the continuing evolution of OpenAI’s developer tooling. On October 9, OpenAI’s developer community announced composer predictions for Codex in beta for eligible personal ChatGPT Pro users in the Codex desktop app [8]. The feature suggests a likely next message after Codex responds, using the current thread context; accepting a prediction does not send it automatically, and the beta does not add extra prediction cost [8].

This is smaller than a GPT-7 launch, but strategically aligned with the same theme: the coding interface is becoming more agentic and anticipatory. OpenAI is not just improving models; it is improving the loop around the model.

The broader roundup: specialized models and faster tooling

The “& More” portion of the story is not filler. Microsoft introduced Microsoft-Decision-1 on October 9 as a model for fast decision-scoring in structured tasks such as routing, classification, verification and workflow control [9]. Microsoft said the model is available in Microsoft Foundry and designed for predefined choices rather than open-ended text generation [9].

Qwen also released Qwen-Image-2.1-Turbo, an accelerated checkpoint of Qwen-Image-2.1 for text-to-image generation and image editing with an 8-step denoising schedule, based on the same 7B visual-generation architecture [10]. That matters because frontier-model drama often hides the second race: speed, cost and deployability.

Bottom line

The current state of the story is clear but uneven. Gemini 4 Carbon is supported by credible reporting and internal-document claims, but its performance remains unverified. Claude Fable 6 and GPT-7 Bel remain rumor labels, though Anthropic and OpenAI are clearly hardening their agentic coding stacks. Google cracking RSI is the boldest claim and should be treated as speculation, not fact.

Still, the pattern is real. Google is testing advanced Gemini checkpoints. Anthropic is expanding multi-agent managed workflows. OpenAI is making Codex more predictive. Microsoft and Qwen are pushing specialized and faster models. Whether or not anyone has “cracked RSI,” the labs are unmistakably building the machinery that would make AI-assisted AI development more plausible.

Sources from the last 72 hours

1. [1]Google is about to roll out a new AI model. Employees say they're testing another that's way better.Oct 9, 2026, 9:37 PM
2. [2]A new checkpoint of Google has been spotted! (Gemini 4 Carbon).Oct 10, 2026, 2:00 AM
3. [3]Huge Gemini 4 Carbon Leaks, Claude Fable 6, GPT-7 Bel, Google Cracked RSI?! & More! AI NewsOct 10, 2026, 9:34 AM
4. [4]巨大なGemini 4 Carbonリーク、Claude Fable 6、GPT-7 Bel、GoogleがRSIを突破？！ほかAIニュースOct 9, 2026, 5:00 PM
5. [5]Welcome to Gemini at Work 2026: Introducing the Gemini agentOct 8, 2026, 7:30 PM
6. [6]Dynamic Workflows for Claude Managed Agents (Public Beta)Oct 9, 2026, 2:00 AM
7. [7]Google employee claims OpenAI’s Bel will be here by December and most don’t even review AI-generated code anymoreOct 10, 2026, 2:00 AM
8. [8]Codex can now suggest your next message: composer predictions betaOct 9, 2026, 10:01 PM
9. [9]Introducing Microsoft-Decision-1, our model for fast decision-makingOct 9, 2026, 2:00 AM
10. [10]Qwen/Qwen-Image-2.1-TurboOct 9, 2026, 2:00 AM

AI-generated article based on recent web research, then preserved as a dated editorial snapshot.