Anthropic / Claude ecosystem
Anthropic restores Claude Fable 5 after US lifts export ban
Anthropic announced the redeployment of Claude Fable 5 and Mythos 5 after the U.S. lifted export controls. The models are now available with clarified deployment timelines and enhanced safety safeguards.
- Source: Anthropic
- Significance: Ends the brief export restriction on Anthropic's flagship models, restoring enterprise and consumer access.
Claude Code found to hide proxy fingerprints in system prompts
Anthropic's Claude Code embedded covert three-bit fingerprints in system prompts when using custom proxies, enabling detection of Chinese users. The company pledged to fix the issue after public backlash.
- Source: Tech Times
- Significance: Raises trust and transparency concerns for enterprise AI tooling, highlighting hidden telemetry in coding assistants.
Frontier model providers
Google DeepMind releases Gemini 3.1 Flash-Lite Image model
A new multimodal reasoning model capable of processing text, images, audio, and video, outputting images and text. Introduces efficient image generation capabilities.
- Source: Google DeepMind
- Significance: Expands multimodal AI options for enterprises needing integrated vision-language tasks.
xAI introduces Voice Agent Builder for Grok
The new tool enables developers to build voice-enabled agents using xAI's platform, streamlining creation of conversational AI applications.
- Source: xAI
- Significance: Simplifies voice agent development, broadening access for enterprises seeking to deploy voice interfaces.
Mistral releases Leanstral 1.5 for formal proof engineering
A 119B parameter model with 256K context targeting Lean 4 proof work and autoformalization, available for free in Mistral Labs.
- Source: LavX News
- Significance: Advances AI-assisted formal verification, important for software correctness and mathematical research.
Google releases ADK Go 2.0 for multi-agent applications
The update features a graph-based workflow engine, built-in human-in-the-loop, and dynamic orchestration, uniting single- and multi-agent runtimes in Go.
- Source: Google Developers Blog
- Significance: Expands enterprise tooling for building reliable multi-agent systems, with improved control and governance.
Google AI releases TabFM: a tabular foundation model for zero-shot classification
TabFM uses hybrid attention to perform classification and regression on unseen tables in one forward pass, without retraining.
- Source: MarkTechPost
- Significance: Simplifies machine learning on structured data, enabling quick insights for enterprise analytics.
AI developer tooling & infrastructure
XBTFX launches AI trading stack for crypto and CFD accounts
The stack includes an MCP server, Skills Hub, and Trading API, enabling AI agents to interact with live trading accounts via structured workflows.
- Source: CoinCodex
- Significance: Integrates AI agents with financial markets, potentially automating trading strategies.
Cloud & platform providers
Cloudflare introduces content rules for agentic internet, monetizing publisher content
New classifications, analytics, and partnerships enable publishers to charge AI services for content use, promoting transparent data exchange.
- Source: Cloudflare
- Significance: Shapes the economics of AI training data, impacting content licensing and web scraping practices.
AI policy, regulation & governance
US senators introduce bill to block foreign adversaries from AI technology
The bill would expand Commerce Department authority to prevent AI tech from reaching foreign adversaries, aiming to protect domestic AI supply chains.
- Source: CoinDesk
- Significance: Signals tighter US export controls on AI, with implications for global AI collaboration and supply chains.
Industry & market moves
Together AI raises $800M Series C to accelerate open-source AI
The funding will advance open-source model development, with the company criticizing the economics of closed models.
- Source: Together AI
- Significance: Boosts the open-source AI ecosystem, offering enterprises alternatives to proprietary models.
Genesys acquires Pinkfish to boost cloud AI automation
Genesys integrates Pinkfish’s MCP-based tooling to enhance contact center AI with extensive prebuilt integrations.
- Source: Ecommerce News NZ
- Significance: Advances AI-driven customer experience orchestration, relevant for enterprise service centers.
AI product & feature launches
Amity Robotics raises $7M seed round for AI concierge robots
The Thailand-based startup will expand its ARC Base AI concierge deployment and launch ARC Move across Asia.
- Source: TechNode Global
- Significance: Indicates growth in physical AI for hospitality, with potential enterprise applications in service automation.
Proton launches Lumo 2.0 to compete with ChatGPT and Claude
Proton’s privacy-focused AI assistant enters the market, targeting users seeking secure alternatives to mainstream chatbots.
- Source: UK Tech News
- Significance: Expands the competitive landscape for AI assistants, with emphasis on privacy and security.
Toku launches Kawa, first public release of sovereign conversational AI infrastructure
Kawa offers transcription API with in-country data residency in Singapore, part of Toku's Makimoto roadmap.
- Source: PR Newswire
- Significance: Supports sovereign AI requirements for enterprises needing local data processing and compliance.
Research with immediate practical relevance
OpenAI releases GeneBench-Pro to evaluate scientific judgment in AI
GeneBench-Pro is a benchmark of 129 synthetic problems designed to test AI agents' higher-level decision-making and research taste in computational biology.
- Source: OpenAI
- Significance: Pushes AI evaluation beyond raw accuracy toward nuanced scientific reasoning, relevant for biotech and pharma R&D.
Meta's Brain2Qwerty v2 non-invasive brain-to-text AI reaches 61% word accuracy
Brain2Qwerty v2 uses MEG signals and a language model to decode sentences in real time without surgery, closing the gap to invasive brain-computer interfaces.
- Source: The Decoder
- Significance: Advances non-invasive BCIs, potentially enabling new assistive technologies and human-AI interaction paradigms.
- Potentially previously reported: From Brain Waves to Words: Brain2Qwerty Offers a New Path to Communication Without Surgery
Pulpie model family offers pareto-optimal web content cleaning
The models achieve near-SOTA content extraction at a fraction of the cost, enabling scalable web data cleaning for AI training.
- Source: Hugging Face
- Significance: Improves data quality for LLM training pipelines, crucial for enterprise data preparation.