AI
Claude Sonnet 5, DeepSeek speed boost, Meta's teen probe, AI shopping wins Prime Day
Anthropic's new Sonnet model, DeepSeek's efficiency breakthrough, Meta's controversial chatbot testing, and AI referrals beat search on Prime Day.
Anthropic Launches Claude Sonnet 5 with Enhanced Agentic Capabilities
Anthropic introduced Claude Sonnet 5, built to be the most agentic Sonnet model yet, capable of planning, using tools like browsers and terminals, and running autonomously at a level previously requiring larger and more expensive models. Its performance is close to Opus 4.8 but at lower prices, with substantial improvements in reasoning, tool use, coding, and knowledge work. Safety assessments show an overall lower rate of undesirable behaviors than Sonnet 4.6, though it scored slightly higher on misaligned behaviors compared to Opus 4.8 and Claude Mythos Preview. The model was never able to develop a full working exploit for vulnerabilities in Firefox 147. Sonnet 5 is available across all plans starting today, with introductory pricing of $2 per million input tokens and $10 per million output tokens through August 31, 2026, after which prices increase to $3 and $15. It is also available via the Claude API, AWS, and Microsoft Foundry, with increased rate limits across all platforms. Early access partners reported that Sonnet 5 finishes complex tasks where previous Sonnet models would stop, checks its own output without being asked, and handles sustained coding and debugging well.
Introducing Claude Sonnet 5 \ Anthropic →
DeepSeek’s DSpark Boosts AI Inference Speed by Up to 85 Percent
DeepSeek released DSpark, a new method that increases per-user response speed for its AI models by 60 to 85 percent. Most LLMs generate text one token at a time, causing low GPU utilization and long wait times; DSpark uses speculative decoding, where a smaller lightweight model proposes answer candidates that the larger model checks in batches, generating small word groups instead of single tokens. A confidence-based system adjusts verification depth on the fly based on compute load, cutting wasted processing on rejected token proposals. Tests with Google DeepMind’s Gemma and Alibaba’s Qwen show the approach works broadly. The framework and DeepSeek-V4-Pro model, developed with Peking University, are available on Hugging Face and GitHub under the MIT license. DSpark achieves the highest text generation efficiency among alternatives like Eagle3 and DFlash. Faster inference reduces chip requirements and infrastructure costs, offering a strategic advantage for China and potentially the EU under tightening US export controls. DeepSeek stated that DSpark enables performance tiers previously unattainable, shifting the Pareto frontier of its serving system.
Meta Used Contractors Posing as Teens to Probe Rival AI Chatbots
Meta ran a project under the internal name Cannes, managed by contractor Covalen, in which hundreds of contractors created dummy under-18 accounts to send prompts and images to competitors’ chatbots and log the replies in spreadsheets. The effort was active as recently as April 21, 2026, targeting OpenAI’s ChatGPT, Google’s Gemini, and Character.AI, none of whom knew the testing was happening. A single round in August 2025 ran over 45,000 prompts; a reviewed spreadsheet of 3,748 prompts included hundreds dealing with suicide, self-harm, eating disorders, and at least 239 involving sex or romance, many taking the voice of a child in crisis. Meta did not deny the work, calling it responsible industry-standard safety testing, but experts and former contractors raised alarms. Rumman Chowdhury, CEO of Humane Intelligence, called the long-term project a governance gray zone where safety becomes cover for anticompetitive practices. All three targeted firms prohibit such testing in their terms of service, and the US Federal Trade Commission has an open inquiry into AI and child safety covering Meta, OpenAI, and Google. Regulators on both sides of the Atlantic are examining accountability when chatbots interact with minors about self-harm.
Meta paid contractors to pose as teens and probe rival AI chatbots →
AI Chatbot Referrals Outperform Search on Prime Day for the First Time
US shoppers spent $26.4 billion across retail sites during the four-day Prime Day event, according to Adobe Analytics. For the first time, consumers who arrived at a retailer via an AI assistant were the most likely to make a purchase — about 40 percent more likely than those coming from search, email, or social media, as reported by GeekWire. On day one, generative-AI traffic to retailers nearly doubled year over year, up 98.3 percent, and AI-referred shoppers spent 49.9 percent longer on site, viewed 20.5 percent more pages, and added items to baskets at a 33 percent higher rate than traditional traffic. Adobe data shows AI-referred traffic to US retail sites rose 393 percent year over year in Q1, and by March the channel was converting about 42 percent better than non-AI traffic, a dramatic reversal from early 2025 when AI visitors converted roughly 38 percent worse. Marketers are calling this shift GEO (generative engine optimization), the new SEO. However, caveats apply: the data comes from Adobe, which sells both analytics and AI tools, and AI referrals remain a small share of total retail traffic despite rapid growth. The four-day sale is not a full year, but the trend signals a fundamental change in how consumers discover products online.