AI News: July 25, 2026
1. Sakana Says Fugu Ultra v1.1 Router Beats Fable 5
Sakana AI. Sakana AI updated its Fugu Ultra model router to v1.1, claiming gains of up to 7.9 points over v1.0 with the largest improvements on ProgramBench and TerminalBench 2.1, and asserting it outscores Anthropic’s Fable 5 without including that model in its routing pool. Pricing holds at 5 dollars per million input tokens and 30 dollars per million output tokens, and a new Claude Code-compatible endpoint adds terminal access. All figures come from Sakana with no independent verification yet, and the service still excludes the EU and EEA over regulatory concerns. Source
2. Bluesky Turns Its Assistant Attie Into a Research Tool
Bluesky. Bluesky expanded its AI assistant Attie with a beta feature called Quests that turns it into an open research tool for asking open-ended questions about news and conversations across Bluesky and other apps built on the AT Protocol. Users can surface trending topics and identify influential accounts in a given field or region, with the company framing it as a way to counter misinformation. Attie previously helped users build custom feeds, and Quests is rolling out gradually from a waitlist. Source
3. Report Finds ChatGPT Health Advice Depends on Subscription Tier
OpenAI. A report on OpenAI’s new Health in ChatGPT feature found that the quality of medical guidance depends on subscription tier, with free users served by GPT-5.5 Instant and paying subscribers getting GPT-5.6 Sol. On HealthBench Professional the two models showed an 88.0 versus 53.2 percent completeness gap and an 83.0 versus 50.8 percent gap in decision helpfulness, even as roughly 300 million people ask ChatGPT health questions each week. Analysts cautioned that the models still stated incorrect findings confidently, unlike human radiologists who better recognized their own uncertainty. Source
4. Prentis Seeks $100M for Computer-Use Agents
Prentis. Prentis, a computer-use AI lab co-founded by CEO Ritankar Das with Reid Hoffman and Mark Pincus and launched in April 2026, is in talks to raise 100 million dollars at a 1 billion dollar valuation. The company trains agents to operate software and automate office work such as insurance claims and customs refund exceptions, and says its Hive-32B model leads rivals on computer-use benchmarks. It has already signed customer contracts worth up to 50 million dollars. Source
5. Cognition Acquires Poke Maker for Its Personality
Cognition. Cognition, maker of the Devin coding agent, acquired The Interaction Company of California, which builds the messaging assistant Poke, in a deal co-founder Marvin von Hagen said valued the startup in the low nine figures. Cognition wants to fold Poke’s conversational personality into Devin, betting that how an assistant talks to users is becoming as important as the underlying model. Poke operates across iMessage, SMS, Telegram, and WhatsApp. Source
6. Midjourney Buys Astrology App Co-Star
Midjourney. Image and video generator Midjourney acquired Co-Star, a social astrology app with about 4.3 million monthly active users, for undisclosed terms, bringing over its roughly two dozen employees. Midjourney said the deal adds consumer-app expertise it lacks, since its own products mostly run through Discord rather than standalone apps. The move signals Midjourney casting a wider net beyond generative media. Source
7. AI Companies Warn Against Broad Open-Weight Restrictions
AI industry coalition. Hugging Face, Meta, Microsoft, Mistral, Nvidia, and Replit signed an open letter urging US policymakers to avoid premature, sweeping restrictions on open-weight AI models. The signatories argued that distillation is a widely used improvement technique that should not be equated with intellectual property theft, and that defenders need openly available models of comparable capability to counter AI-enabled cyberattacks. The letter responds to the administration’s deliberations over banning Chinese open-weight models amid allegations against labs such as Moonshot. Source
8. EU Publishes Digital Omnibus Amending the AI Act
European Union. The EU’s Digital Omnibus on AI was published in the Official Gazette on July 24, 2026, enacting the first amendments to the AI Act since its 2024 adoption. The changes push compliance deadlines for standalone high-risk systems to December 2, 2027 and for AI embedded in regulated products to August 2, 2028, cut overlapping registration requirements, soften AI literacy obligations, and add prohibitions on non-consensual sexual imagery generation. Officials framed the package as simplification rather than deregulation. Source
9. Google’s Gemini Nears a Billion Users
Google. Google said its Gemini assistant surpassed 950 million monthly active users, putting it on track to join the company’s roster of billion-user products alongside Search, Gmail, Android, and YouTube. Usage has roughly tripled year over year, up from 750 million monthly users reported in February 2026. The growth underscores how quickly consumer AI assistants are scaling. Source