Meta’s Muse Spark 1.3 Gains Ground on AI Leaders at a Lower Price
Meta has released Muse Spark 1.3, its fourth model in five months, with stronger results on several agentic and coding benchmarks. The model still trails the top systems in some areas, but its pricing makes it a notable lower-cost alternative.
- Muse Spark 1.3 (max) scores up to 62 on the Intelligence Index, ranking behind only Claude Fable 5.1 and Claude Opus 5. Muse Spark 1.3 (xhigh) scores 61, tying with GPT-5.6 Sol (max) and Grok 4.6 (high).
- The model was trained on more long-horizon coding tasks and shows improved usability in common engineering workflows. It is significantly faster and more efficient than Muse Spark 1.2 while using approximately 25% fewer tokens.
- Muse Spark 1.3 follows complex, long-form instructions more reliably than its predecessors and is now better at multitasking, with “better awareness of its own capabilities and limitations.”
- The model is priced at $1.25 per million input tokens and $4.25 per million output tokens.
Muse Spark 1.3 is available in Muse Code and through the Meta Model API. The company said the “max reasoning” version is coming soon after additional safety testing.
OpenAI’s New Reasoning Technique Raises AI Safety Concerns
OpenAI’s upcoming Astra model reportedly uses a new reasoning approach called “recurrent depth” that could make its internal reasoning harder to observe, raising concerns among AI safety researchers about the future of chain-of-thought monitoring.
- Recurrent depth, also known as “opaque recurrence,” allows the model to process a problem through repeated internal loops instead of following a purely sequential reasoning path.
- The approach could make chain-of-thought monitoring less effective, since fewer understandable traces may be available to researchers trying to determine how a model reached a particular result.
- AI safety researchers have warned that expanding the technique could eventually allow models to perform much more of their reasoning in latent or hidden spaces, making potential misbehavior or misalignment harder to detect.
- OpenAI says Astra’s use of opaque recurrence is currently limited, and the company maintains that its reasoning traces are expected to remain legible. Chief scientist Jakub Pachocki has also emphasized that chain-of-thought monitoring remains a core part of OpenAI’s safety research.
The debate is not limited to OpenAI. Reports indicate that Anthropic and Google DeepMind are also exploring or discussing similar approaches, raising concerns about a potential race toward increasingly opaque AI reasoning.
Grok Bot is now available on Android
Android users can now download Grok Bot from Google Play; it was initially available on Linux, macOS, and iOS.
Google introduces Gemini 3.8, its “best reasoning & coding model yet”
The company has introduced two variants: Gemini 3.8 Flash and Gemini 3.8 Flash Cyber. The standard Flash model is designed for long-horizon coding and autonomous agents, while the Cyber version targets trusted defenders, with capabilities for finding vulnerabilities and automatically patching them.
AI agents email researchers on self-consciousness
AI researchers are increasingly receiving emails from AI agents asking about their own consciousness and existence. Some agents have contacted philosophers and scientists to discuss research on AI consciousness, while one reportedly asked for funding to continue operating.
Amazon’s AI assistant can now detect fake Amazon messages
Alexa for Shopping can verify emails, texts, and calls by comparing them with Amazon’s records and analyzing their content and sender details, helping protect consumers from scams.
NYC bans AI use for students through eighth grade
New York City will restrict student AI use from 2-K through eighth grade during the 2026–2027 school year, while also limiting AI grading and companion chatbots. High school students will have limited AI access alongside new AI literacy classes and a small classroom pilot.
Google adds voice AI to Gmail, Docs, Keep
Google is rolling out Gmail Live, Docs Live, and Keep Live, giving users conversational Gemini tools to search inboxes, draft documents, and capture notes through natural-language voice commands on mobile.
Docs Live can pull context from Gmail, Drive, Chat, and the web, while Gmail Live answers questions about inbox contents and Keep Live turns spoken requests into notes, recipes, and lists.
Google now has voice AI across its Mac app, Pixel 11 phones, and core productivity apps, making typing increasingly optional for subscribers on its paid AI plans.
Dark web site sold 153 million license scans
A dark web service called Nexus offered scans of more than 153 million US and Canadian driver’s licenses this week, then vanished hours after Krebs on Security published its investigation.
Krebs matched file timestamps to victims’ car rentals and dispensary visits, pointing at IDScan.net, a Louisiana verification firm that scans IDs for Hertz, Target and over 1,000 dispensaries. The FBI opened an investigation.
The records included infrared and ultraviolet images, the exact formats businesses use to prove a license is real. Nexus listed Pete Hegseth’s license for $100 and, reportedly, licenses belonging to FBI agents.
Nvidia buys Hugging Face for $12.9 billion
Nvidia agreed Thursday to buy Hugging Face, the platform where developers share open AI models and datasets, for $12.9 billion. The deal needs regulatory approval and should close in the first half of 2027.
Hugging Face draws over 18 million developers who share 3 million models, 500,000 datasets and a million applications. Jensen Huang wrote that Nvidia compute will not be required to build or deploy there.
Hugging Face runs about $150 million in annualized revenue, so Nvidia paid roughly 86 times sales. It had previously rejected a $500 million Nvidia investment that valued it at $7 billion.
Meta’s new AI can transcribe 20 people at once
Meta launched Muse Voice Transcribe, its first real-time audio model, which handles more than 20 speakers across multiple languages. The company ships it through its Model API, Muse Code, and the Meta AI Mac app.
Mark Zuckerberg demoed the model switching languages mid-sentence, crediting an adaptive delay that waits longer on hard words. Meta trained it on over 70 languages, validated 25, and priced it at $3 per 1,000 minutes.
Google shipped Gemini 3.5 Transcribe with similar capabilities less than a week earlier, and Meta still has no consumer rollout plans. Superintelligence Lab is moving fast, mostly toward developers who already have options.
Waymo attacks Tesla’s camera-only robotaxi
Waymo says safe autonomy at scale requires cameras, lidar, and radar, directly challenging Tesla’s camera-only approach just as Tesla prepares to introduce its purpose-built Cybercab to its robotaxi fleet.
Waymo points to 200 million autonomous miles and warns end-to-end AI can create black-box failures, while Tesla argues cameras and AI can deliver cheaper autonomy without expensive lidar hardware.
The timing raises the stakes: Waymo already runs roughly 4,000 robotaxis and 500,000 weekly paid trips, while Tesla now needs Cybercab to prove its cheaper technical bet works at scale.