Tech • AI • Robotics • Game
Claude OPUS 6 EAP SPOTTED?! Muse Spark Open, GPT-6.1 Astra, MAI-Base-2, & Fable 5.5 LEAKS! AI NEWS
TL;DR
Meta has released Muse Spark Open, an unusually large open-weight multimodal model aimed at coding and agentic work, as reports also point to fresh model activity from OpenAI, Microsoft, Anthropic, Nvidia, Google, and Tesla.
KEY POINTS
Meta unveils Muse Spark Open
Muse Spark Open is positioned as a major open-weight release, with 816 billion total parameters, 42 billion active, and a 1 million-token context window. It is natively multimodal across text, images, video, and audio, and is targeted at coding, AI agents, software engineering, and long-horizon reasoning. The model weights are available publicly, though access is gated by identity details.
Frontier-scale model can run locally
The most striking claim is local deployment. Reports indicate the model can run on very high-end desktop-class hardware, including configurations built around Apple Ultra systems with hundreds of gigabytes of unified memory. That does not make it consumer-friendly, but it narrows the gap between cloud-only frontier systems and locally controlled AI.
Coding performance appears competitive
In one coding comparison centered on building a Vampire Survivors-style browser game with HTML canvas, Muse Spark Open reportedly produced the most playable result among four open-weight models. The output included enemy behavior, damage numbers, a level-up loop, and no crashes, finishing in about 20 minutes at an estimated cost of $0.59 on hosted inference. It was not the fastest, but it delivered the strongest overall functionality in that test.
Limits remain in visual composition
Tests involving 3D and voxel-style generation suggest strong reasoning but uneven visual fidelity. In a pagoda-building benchmark, the model successfully created a five-tier structure with a gate on the first attempt, but the scene was cropped and relatively sparse in detail. That points to a recurring tradeoff: strong code and structural reasoning, weaker taste and precision in final visual composition.
Anthropic model clues deepen
New checkpoint names linked to Anthropic have fueled speculation around upcoming releases beyond Opus 5.5 Fast. Labels including Claude Oni Giddy EAP, Claude Oniri EAP, and Claude Gabagul EAP have been associated with multiple reasoning settings and possible fallback behavior to Opus 4.8. The working theory is that one checkpoint may map to Fable 5.5, while another could point to an Opus 6-class system.
Opus 5.5 Fast broadens access
Anthropic has officially moved Opus 5.5 Fast out of research preview and removed its waitlist. The model is marketed as up to 2.5 times faster, with pricing at $8 per million input tokens and $40 per million output tokens. It is also expanding through major cloud platforms, while Claude Code gains practical workflow features such as session export.
OpenAI activity points to GPT-6 Astra
A newer internal snapshot of GPT-6 Astra has reportedly surfaced, suggesting another release may be approaching under OpenAI’s accelerated product cadence. The company has also pushed service-side updates to reduce “model at capacity” errors and improved auto-resume behavior for interrupted sessions. Voice features are also expanding with broader language support and a newer live model.
Microsoft trains a giant MAI Base 2
Microsoft is said to be finishing pre-training on MAI Base 2, a 2.8 trillion-parameter model with 80 billion active parameters. The system was reportedly trained on 16,000 Nvidia GB200 GPUs in 38 days and is described as multimodal. Early claims say it outperforms nearly every tested Chinese base model, though post-training results will determine its real competitiveness.
Free and open model competition intensifies
DeepSeek 4.1 Flash has reportedly been made temporarily free in some coding tools, bringing a 200,000-token context window to no-cost users. Nvidia has also introduced Nemotron 3.5 SuperVL, an open model with 120 billion total parameters and 12 billion active, built for coding, reasoning, and visual understanding in one stack. Black Forest Labs added Flux 3 Kontext, a 5.6 billion-parameter image model designed for local use and native 4K generation.
Google and Tesla add to the week’s pace
Google is rolling out Gemini 4 Argon more broadly to paid subscribers with a separate weekly quota for heavier reasoning tasks. Outside core AI software, Tesla demonstrated a new Roadster performing airborne maneuvers using compressed-air thrusters developed with SpaceX, highlighting how advanced control systems and AI-assisted real-time coordination are spreading into robotics-adjacent products.
CONCLUSION
The latest wave of releases and leaks shows the AI race expanding on two fronts at once: larger frontier systems from major labs and more capable open-weight models that users can increasingly control themselves. Meta’s Muse Spark Open is the clearest sign yet that local AI is moving closer to frontier-class performance, even if the required hardware remains extreme.
Ask a question