AI News Top 10 — June 28, 2026
This week’s AI landscape is defined by major model launches, strategic hardware plays, and geopolitical shifts. Here are the top 10 developments you need to know.

The Top 10
- OpenAI Launches GPT-5.6 “Sol” — Three-Layer Model Family
On June 26, OpenAI previewed GPT-5.6, a three-tier model family codenamed “Sol.” Building on GPT-5.3 Codex-Spark, it promises significant improvements in reasoning and agentic tasks. The release also raises government questions around safety and export controls.
Context: This is OpenAI’s answer to growing competitive pressure from Anthropic and open-weight models. The three-tier structure (likely a smaller/faster variant alongside the flagship) suggests a strategy to serve different price-performance segments simultaneously.
Source: openai.com · alphamatch.ai - OpenAI & Broadcom Unveil “Jalapeño” — Custom AI Inference Chip
On June 24, OpenAI and Broadcom revealed the Jalapeño chip, OpenAI’s first custom AI accelerator designed specifically for LLM inference. Developed in just 9 months — the fastest ASIC cycle ever — and manufactured by TSMC, early tests show ~50% cost savings vs. NVIDIA GPUs. Engineering samples are already running GPT-5.3 Codex-Spark at production frequency.
Context: This is a strategic play for independence from NVIDIA, whose GPUs currently dominate AI inference. A 50% cost reduction at scale would fundamentally reshape the economics of running large models. Expected in Microsoft datacenters by end of 2026.
Source: openai.com · iex.nl - Anthropic’s “Claude Tag” — Persistent AI Teammate in Slack
On June 23, Anthropic launched Claude Tag, an always-on @Claude teammate in Slack that reads channel context, delegates multi-step work asynchronously, learns from conversations, and proactively shares updates. Available in beta for Claude Enterprise and Team customers, powered by Opus 4.8.
Context: This creates a new product category: “AI as a colleague” embedded in enterprise communication. Unlike chatbots that require explicit prompting, Claude Tag proactively engages — a significant step toward truly autonomous AI agents in the workplace.
Source: anthropic.com · mer.vin - Apple Rebuilds Siri with Google Gemini — $1B/Year Deal
At WWDC 2026, Apple announced that Siri is being completely rebuilt on a custom Google Gemini model. A multi-year deal worth approximately $1 billion per year. The new Siri AI rolls out with iOS 26.4 and iOS 27, marking a historic strategic shift: Apple is handing the “brains” of its assistant to a competitor.
Context: This is one of the largest AI licensing deals ever. For Apple, it acknowledges that building a competitive LLM from scratch wasn’t practical. For Google, it’s a massive validation of Gemini’s capabilities — and a revenue stream that could eclipse search ads for AI.
Source: apple.com · kersai.com - NVIDIA Nemotron 3 Ultra — Most Powerful Open-Weight US Model
Jensen Huang unveiled the Nemotron 3 Ultra at Computex 2026: 550B total parameters, 55B active per token via MoE, 300+ tokens/sec, 1M token context window, and an Intelligence Index score of 48 — the best-performing open model from the US. Open weights on Hugging Face with published training recipes.
Context: This sets a new bar for open-source AI capabilities. The combination of massive context window and high throughput makes it competitive with closed models, while published training recipes enable the research community to build on NVIDIA’s work.
Source: research.nvidia.com · buildfastwithai.com - Anthropic IPO Shaken by National Security Issue with Trump
Anthropic confidentially filed for an IPO in fall 2026, but the Trump administration restricted foreign access to Anthropic’s latest Fable 5 and Mythos 5 models over national security concerns. After a meeting between Trump and CEO Dario Amodei at the G7 summit, Trump declared Anthropic “no longer a national security threat.” The company is now allowed to release Mythos to select businesses and agencies.
Context: This sets a precedent for AI export controls that directly impacts IPO valuation. The G7-level intervention underscores how AI companies are increasingly subject to geopolitical dynamics. The resolution suggests a model where top AI firms negotiate directly with governments on release policies.
Source: bloomberg.com · reuters.com - Stanford AI Index 2026: US vs China Neck-and-Neck, Investments 100× Manhattan Project
The Stanford HAI AI Index Report 2026 shows US and Chinese models have repeatedly traded the lead since early 2025. As of March 2026, Anthropic’s top model leads by just 2.7%. AI investments are now 100 times larger than the Manhattan Project (inflation-adjusted). The intelligence gap between top models is rapidly shrinking.
Context: This report provides critical evidence for global AI strategy debates. The 100× Manhattan Project comparison puts AI spending in historical perspective, while the narrowing gap between US and Chinese models suggests the competitive window is closing.
Source: hai.stanford.edu - Google Gemini Expands: Live Translate and Deeper Integration
Google launched real-time Gemini Live Translate with unprecedented accuracy across multiple languages, plus deeper Gemini integration into Google Calendar, Gmail, and Workspace. The assistant can now perform context-aware actions like managing appointments and drafting emails based on user preferences.
Context: Google is embedding AI into every layer of its product ecosystem. Live Translate positions Gemini against dedicated translation tools, while Workspace integration turns Gemini into an always-on productivity layer for billions of users.
Source: imfounder.com · kersai.com - Qualcomm Forecasts $15B Datacenter Chip Revenue by 2029
Qualcomm announced it expects to generate $15 billion in revenue from datacenter activities by 2029, expanding beyond its core smartphone chip market. The stock rose sharply on the news, underscoring growing demand for AI inference hardware outside NVIDIA’s ecosystem.
Context: Qualcomm’s push into datacenter AI signals that the inference hardware market is diversifying. Combined with OpenAI’s Jalapeño chip and NVIDIA’s dominance, we’re seeing the early stages of a multi-vendor AI chip landscape.
Source: iex.nl · reuters.com - Open-Source AI Summer: MiniMax M3, NVIDIA Cosmos 3, New Agent Toolkits
June 2026 is seeing a wave of open-source AI releases: MiniMax M3, NVIDIA Cosmos 3 (physical world models), OpenClaw updates, and several new agent toolkits for developers. The open-source community is building increasingly capable alternatives to closed-source models, especially for robotics and multi-agent systems.
Context: This open-source momentum is democratizing AI development. Physical world models (Cosmos 3) and multi-agent frameworks are particularly notable — they enable capabilities that closed-source labs are only beginning to explore.
Source: devflokers.com · osasai.com
Key Trends This Month
- Hardware independence is accelerating. OpenAI’s Jalapeño chip and Qualcomm’s datacenter push show the industry is actively building alternatives to NVIDIA’s near-monopoly — expect more custom silicon announcements in H2 2026.
- Geopolitics shapes AI access. The Anthropic IPO / Trump administration saga demonstrates that national security concerns now directly influence which models reach the market and who can use them. Export controls are the new normal.
- AI is becoming a co-worker, not just a tool. Claude Tag in Slack points toward a future where AI agents participate in workflows proactively — reading context, delegating tasks, and reporting back without being asked.
- The US-China AI gap is nearly closed. Stanford’s report confirms that Chinese models are within 3% of US leaders. The race is no longer about who leads, but who deploys faster and at scale.
- Open-source capability is surging. With Nemotron 3 Ultra and Cosmos 3, open-weight models now rival top closed models in specific domains. The “open-source AI summer” is real and accelerating.
- Enterprise AI integration hits new highs. Apple rebuilding Siri on Gemini and Google embedding Gemini into Workspace mark a phase shift: AI is becoming the operating system layer of enterprise software.
Wat vond je van dit bericht?