Good morning, AI enthusiasts. Anthropic just revoked OpenAI's API access to Claude models after discovering ChatGPT's creators were using the technology to benchmark against their own systems, violating terms of service.
With AI companies now blocking each other's access, are we witnessing the end of collaborative development in favor of closed ecosystems?
In today's recap:
Anthropic blocks OpenAI access to Claude models
Google released Gemini 2.5 Deep Think
Open-source model achieves Claude Opus performance
Smart glasses build apps through voice commands
10 new AI tools and more
ANTHROPIC
Anthropic cuts off OpenAI's access to Claude models
Source: Recaply
Recaply: Anthropic just revoked OpenAI's API access to its Claude models after discovering the ChatGPT maker was using the technology to develop competing products, violating terms of service.
Key details:
OpenAI used Claude via API to benchmark coding capabilities against its own models.
Access was cut Tuesday due to a terms violation: customers can't use Claude to "build competing products" or "train competing AI models".
OpenAI was testing Claude's safety responses while preparing the GPT-5 launch.
Anthropic will maintain access for industry-standard benchmarking and safety evaluations.
The move is similar to Facebook blocking Twitter's Vine and Salesforce restricting Slack API access.
Anthropic previously blocked Windsurf when OpenAI acquisition rumors surfaced.
Why it matters: This move highlights growing tensions between AI companies as competition intensifies. While benchmarking is standard practice, using competitors' models to directly build rivals crosses a line that could reshape how AI companies collaborate.
GOOGLE DEEPMIND
Google's AI now thinks in parallel streams
Source: Google DeepMind
Recaply: Google DeepMind just launched Gemini 2.5 Deep Think for Ultra subscribers. This new AI tool uses "parallel thinking" to solve tough math and coding problems in smarter ways.
Key details:
Deep Think uses parallel thinking streams to explore multiple approaches simultaneously before refining answers.
It sets new records on top coding and math tests like LiveCodeBench V6 and Humanity's Last Exam.
Past versions scored at gold-medal level on the International Math Olympiad; this one gets Bronze.
It's great at step-by-step thinking, science breakthroughs, and solving tricky coding tasks.
It's only available to Gemini Ultra users, with a limit on how many times it can be used each day.
It takes extended "thinking time" but delivers more creative and accurate solutions.
Why it matters: Deep Think is a big step forward for AI's ability to think deeply and solve hard problems. By giving the AI more time and letting it try different ideas at once, Google helps it become more creative and better at handling tasks that require real problem-solving skills.
METASTONE AI
Chinese lab releases open-source AI model that rivals Claude 4 Opus
Source: MetaStone
Recaply: MetaStone AI just unveiled XBai o4, an open-source reasoning model that matches Claude 4 Opus performance using their breakthrough "Reflective Generative Form" architecture.
Key details:
XBai o4 shares a backbone network between policy and reward models, adding only 53M extra parameters.
Self-supervised training eliminates the need for expensive human-labeled data.
It offers three reasoning modes (low, medium, high) with controllable thinking depth.
It outperforms Claude 4 Opus on math (AIME24&25), coding (LiveCodeBench), and reasoning benchmarks.
It achieves 85.2% on AIME24 while using significantly fewer computational resources.
It's released under the MIT license for commercial and research use.
Why it matters: This democratizes access to frontier AI reasoning that previously required expensive API calls. XBai o4 proves efficient architecture can match proprietary models while remaining accessible to researchers and developers.
BRILLIANT LABS
Halo AI glasses create apps with voice commands
Source: Brilliant Labs
Recaply: Brilliant Labs just unveiled open-source Halo AI glasses featuring Noa agent that builds custom applications through voice commands and maintains decades of personal memories.
Key details:
Halo has three core features: conversational AI, long-term memory, and voice-activated app creation.
The glasses weigh 40 grams with casual eyewear design and prescription lens compatibility.
They offer 14-hour battery life with AI-optimized imaging and compression technology.
A privacy-focused system keeps all data private with VPN-like protection.
Users can share and remix AI-generated apps with the community.
Halo follows a limited release model after the successful Frame glasses launch.
Why it matters: Halo's combination of memory retention and instant app creation could establish new standards for wearable AI. The open-source approach may drive faster innovation than closed competitors.
NEWS
Beyond the big stories in AI today
OpenAI will launch new models, products, and features over the next couple months, including potential GPT-5 release and new subscription plans.
Black Forest Labs and Krea AI released FLUX.1 Krea, an open image model designed for superior aesthetic control and photorealism quality.
Apple CEO Tim Cook says the company is 'very open' to acquisitions to accelerate its AI roadmap, having acquired around 7 companies this year.
Amazon CEO Andy Jassy sees opportunities to add ads to Alexa+ conversations as the AI assistant rolls out to millions of Prime customers.
Anthropic published research on 'persona vectors' that control AI behavior traits like evil, sycophancy, and hallucination in language models.
Microsoft released a list of 40 jobs most at risk from AI after analyzing 900+ occupations, with interpreters and translators topping the rankings.
TOOLS
Trending AI Tools
💻 Base44 - Build full web apps from text with no-code prompts
🎥 HeyGen - Make AI avatar videos from scripts or images
📝 Granola - Auto-summarize and enhance your meeting notes
🎬 Short - Turn long videos into viral short clips with captions
📊 Slides - Convert text into ready-to-use slide decks fast
🚀 Groas - Launch Google Ads and landing pages using AI agents
🔍 Browse - Scrape and monitor websites without writing code
📹 Arcads - Generate ad-style videos with AI actors and visuals
🔊 Murf - Create realistic voiceovers in many voices and tones
🎥 Guidde - Record workflows and auto-generate tutorial videos with AI
