Tech • AI • Robotics • Game
Huge Gemini 4 Carbon Leaks, Claude Fable 6, GPT-7 Bel, Google Cracked RSI?! & More! AI News
9.4/10TL;DR
Unverified claims suggest Google, Anthropic, and OpenAI are testing more advanced AI systems, while a wave of new tools and AI-assisted engineering projects points to a faster, more competitive phase in the AI race.
KEY POINTS
Google tests another Gemini checkpoint
Google is reportedly testing an internal Gemini 4 checkpoint codenamed Carbon, even before Gemini 4 Argon has launched publicly. Insider claims say some employees view Carbon as comparable to Claude Opus 5.5 on coding tasks, though no independent benchmark has confirmed that comparison. Reports also mention earlier internal checkpoints such as Beryllium, suggesting Google is iterating several versions in parallel.
Recursive self-improvement rumors intensify
Separate insider claims say Google may be close to or already using forms of recursive self-improvement, in which AI helps accelerate the development of future AI systems. The speculation is being linked to the rapid pace of the Gemini Flash line and the apparent speed of internal model iteration. None of the claims have been independently verified, but they have fueled broader debate about whether major labs are entering a new acceleration phase.
Anthropic and OpenAI face similar speculation
Anthropic is rumored to have started a new pre-training run that goes beyond the expected Claude Fable 5.5, potentially aimed at a more substantial next-generation model. At the same time, insiders claim OpenAI is developing a model known as Bell, which some speculate may underpin recently published mathematical research results. Those claims also include suggestions that AI-assisted research may already be contributing to model development at the frontier labs.
Google expands reasoning and developer modes
New reasoning-related options have reportedly appeared inside Google AI Studio, including an ultra mode alongside plan and build and a security review mode. The additions suggest Google is preparing more advanced workflows for coding and reasoning use cases, potentially aligned with a future Argon release. The move fits a broader trend of AI companies packaging models into more task-specific tooling rather than offering only general chat interfaces.
Qwen releases a faster image model
Qwen Image 2.1 Turbo has been released as a faster version of the earlier 2.1 image model. The system reportedly generates or edits images in 8 denoising steps, down from about 40, an 80% reduction while aiming to preserve similar quality. Built on a 7 billion-parameter architecture, it supports 2K image generation, natural-language editing, local deployment through downloadable weights, and hosted API access.
Microsoft targets structured decisions
Microsoft has introduced Decision 1, a model optimized for fast, structured decision-making rather than open-ended text generation. The company says it performs well on latency and decision quality in evaluated tasks and is being tested for incident response, quality control, and scientific discovery. The release highlights a growing push toward smaller or more specialized models tailored to specific enterprise workflows.
OpenAI and Anthropic add coding workflow upgrades
OpenAI has added Composer Predictions for Codex Pro users, letting the system suggest likely next prompts from project context and prior interactions. It also introduced a Windows-focused sandbox based on Microsoft execution containers to isolate AI agents with tighter file and network controls. Anthropic, meanwhile, launched Claude managed agents dynamic workflows, where multiple agents can divide tasks and recombine results; in one internal test on a 116,000-line codebase with 70 planted bugs, a multi-agent setup found 66, far above a single-agent baseline.
Free and high-efficiency models broaden access
A free model called Solar Mini 4 from Upstage AI has also drawn attention. It uses a 35 billion-parameter mixture-of-experts architecture while activating about 3 billion parameters at a time, supports a 524,000-token context window, and is aimed more at document parsing and processing than coding. The release reflects how labs are increasingly competing not only on peak capability but also on cost, speed, and deployment flexibility.
AI-assisted software engineering reaches unusual targets
Two experimental projects show how AI coding tools are being used on difficult systems work. One developer demonstrated early NVIDIA RTX support on Intel-based Mac and Hackintosh systems running modern macOS, using AI assistance to help build drivers and compatibility layers. Another project, AnyPS5, is exploring PlayStation 5 software compatibility on PC by translating console libraries and graphics behavior, showing how AI-assisted coding is beginning to influence reverse engineering and low-level platform work.
CONCLUSION
Most of the biggest claims remain unverified, especially around frontier model performance and recursive self-improvement. Even so, the volume of new releases, internal testing rumors, and AI-assisted engineering breakthroughs indicates that competition among major AI labs is intensifying rapidly.
Ask a question