- MCP goes stateless today: The Model Context Protocol's
2026-07-28specification lands, removing session management and the initialize handshake so any request can hit any server instance behind a plain load balancer — the biggest change to the agent-tooling standard since launch. - NVIDIA's half-trillion-dollar OpenAI bet: Reports say NVIDIA is in talks to guarantee up to roughly $250B toward an OpenAI data-center lease worth about $500B — tying GPU demand directly to the compute buildout as Vera Rubin ramps across the major clouds.
- Open weights keep surging: Moonshot's ~2.8T-parameter Kimi K3 ships openly on Hugging Face alongside Thinking Machines' Inkling and DeepSeek V4, pushing flagship-scale models into download-and-run territory.
- The frontier stack holds: Claude Opus 5 ($5/$25, ~3x the next-best on ARC-AGI 3), OpenAI's GPT-5.6 family (Sol/Terra/Luna, from $5/$30 to $1/$6), and Google's Gemini 3.6 Flash ($1.50/$7.50) continue to reset price-performance.
AI News (Top Updates)
1. The Model Context Protocol goes stateless
The MCP 2026-07-28 specification — a release candidate since May and finalizing on July 28 — removes session management and the initialize handshake, making the protocol core stateless so any request can hit any server instance behind a plain round-robin load balancer. It adds an Extensions framework (including official "MCP Apps" server-rendered UIs in sandboxed iframes and a redesigned Tasks lifecycle), OAuth/OIDC authorization hardening, and full JSON Schema 2020-12, while deprecating Roots, Sampling, and Logging with 12-month removal windows. Beta SDKs shipped ahead of the release.
2. NVIDIA reportedly in talks to backstop a $500B OpenAI data center
Reports on July 26–27 (WSJ, Bloomberg, Forbes) say NVIDIA is discussing guaranteeing up to roughly $250B of financing toward an OpenAI mega-datacenter lease valued at around $500B — an extraordinary vendor-financing arrangement that would tie GPU demand directly to compute buildout. It lands as NVIDIA's next-generation Vera Rubin platform enters production ramp across CoreWeave, Google Cloud, Microsoft Azure, and Oracle Cloud Infrastructure.
3. Anthropic's Claude Opus 5 leads the frontier
Released July 24, Claude Opus 5 (claude-opus-5) brings major gains in reasoning, self-verification, coding, knowledge work, and science, with a new effort-control toggle that trades intelligence for cost and speed. Anthropic reports it more than doubles Claude Opus 4.8 on its internal Frontier-Bench v0.1 at lower cost, scores roughly three times the next-best model on ARC-AGI 3, and beats its prior top model on OSWorld 2.0 at about a third of the cost. Pricing stays flat versus Opus 4.8 at $5 per million input tokens and $25 per million output, with a Fast mode running ~2.5x faster for double the base price. It is now the default on Claude Max.
4. OpenAI's GPT-5.6 family resets the coding and agents frontier
Announced July 9, the GPT-5.6 family spans three tiers — Sol (flagship), Terra (balanced), and Luna (fastest) — across ChatGPT, Codex, and the API, with a new "ultra" mode that coordinates four agents in parallel by default. Sol scores 80 on the Artificial Analysis Coding Agent Index while using under half the output tokens, and posts a state-of-the-art 92.2% on BrowseComp with up to 1M-token context. New programmatic tool calling in the Responses API cut total tokens by up to 63.5% in testing. Pricing per million tokens runs Sol $5/$30, Terra $2.50/$15, and Luna $1/$6.
5. Google refreshes the Gemini Flash tier and starts training Gemini 4
On July 21 Google shipped Gemini 3.6 Flash — the new default workhorse, cutting output tokens about 17% while improving coding (DeepSWE 49%, up from 37%) at $1.50/$7.50 per million — alongside Gemini 3.5 Flash-Lite ($0.30/$2.50 for high-throughput document and agentic-search work) and Gemini 3.5 Flash Cyber, a security-tuned model paired with CodeMender and limited to governments and trusted partners. Google also confirmed it has begun "the most ambitious pre-training run yet, for Gemini 4," and advanced 3.6 Flash's knowledge cutoff to March 2026.
6. xAI ships Grok 4.5 and pushes into Office workflows
Grok 4.5, an MoE model built for coding and agents, moved from private beta to public availability around July 8–9, posting 83.3% on Terminal-Bench 2.1. xAI followed with a wave of product moves: Grok add-ins for Excel, Outlook, Word, and PowerPoint (July 20–21), Automations for scheduled and email-triggered recurring tasks (July 16), and the open-sourcing of its Grok Build coding-agent framework on GitHub (July 15) with a fully local-first setup.
7. Moonshot opens Kimi K3 weights at ~2.8 trillion parameters
Moonshot AI released Kimi K3 open weights on its official Hugging Face org around July 27, in the ~2.8-trillion-parameter class with MXFP4 quantization and a roughly 1.4TB download — one of the largest openly released models to date, distributed under a revenue-tiered license. It intensifies the open-weight race out of China alongside DeepSeek and Alibaba's Qwen line, letting teams self-host rather than send data to an API.
8. Thinking Machines ships Inkling, a trillion-parameter open model
On July 15 Thinking Machines released Inkling on Hugging Face — a roughly 1-trillion-parameter (about 41B active) decoder-only mixture-of-experts multimodal model handling text, image, and audio, with a 1-million-token context window. Trained on about 45 trillion tokens, it ships in BF16 (≈2TB VRAM) and quantized NVFP4 (≈600GB) formats with day-zero support in transformers, vLLM, SGLang, and llama.cpp; Unsloth 1-bit builds cut memory demands ~95%.
9. DeepSeek V4 lands with first-ever peak-time pricing
DeepSeek confirmed its V4 flagship for mid-July, making a long-context MoE a low-cost, frontier-class option compatible with both OpenAI Chat Completions and Anthropic-style APIs. Alongside it, DeepSeek introduced its first time-based API surcharging: prices double during peak windows (9AM–12PM and 2PM–6PM daily) while off-peak pricing stays the same — a novel way to manage demand on cheap inference.
10. Meta ships Muse Spark 1.1, Muse Image, and Muse Video
Meta Superintelligence Labs launched its first paid Meta Model API on July 9 with Muse Spark 1.1 — a 1M-token agentic model with desktop, browser, and mobile computer use, parallel tool calling, and built-in search with citations, priced around $1.25/$4.25 per million tokens (US-only, $20 free credits). Days earlier it debuted Muse Image (its first in-house image generator, #2 on the text-to-image Arena at launch) and Muse Video (text-to-video with native audio, #3 on the video Arena), both available in Meta AI, Instagram Stories, and WhatsApp.
11. Ollama turns local models into tool-using agents
Ollama's v0.32 line (July 14–27) added an interactive "Chat, Code & Work" agent experience launched from the command line, with web search, task delegation, a new skills system, and unlimited tool rounds for cloud models by default. Later point releases (v0.32.3–.4) expanded GPU support — CUDA on Windows ARM64 and NVIDIA B200 — added Laguna 2.1 model support, and sped up Qwen3 MoE decoding roughly 4–9% on Apple M5 Max.
12. Microsoft wires MCP agents into Office and adds Claude to Copilot
Microsoft 365 Copilot's July release notes bring Model Context Protocol agents directly into Word, Excel, PowerPoint, Outlook, and Catalyst (July 15), plus a governed Agent Store submission flow, AI-content watermarks, and long-running agent-task status in the Windows taskbar. Federated Copilot Connectors — real-time third-party data over MCP with runtime user-level auth — reached general availability, and Anthropic's Claude is now selectable inside Copilot Chat for complex analysis and document work.
13. Ecosystem: Anthropic's open-weight stance, Mistral's physical AI, and media-gen drops
Anthropic expanded its enterprise partnership with Cognizant, published a position statement on open-weight models, and launched an Economic Index connector (July 22–27). Mistral pushed into physical AI with Robostral Navigate — an 8B model steering robots from natural language using a single RGB camera (76.6% on R2R-CE) — plus the Apache-2.0 Leanstral 1.5 for formal math. On media generation, Alibaba shipped Qwen-Image-3.0 (July 20) for photorealism, ElevenLabs' Music v2 added mid-track genre switching and licensed commercial use, and Runway Gen-4.5 integrated with ElevenLabs for professional video.
Top 5 New / Popular AI Products
1. Claude Opus 5
New Flagship Model
Anthropic's new frontier model (claude-opus-5), delivering top-tier reasoning and agentic coding at flat $5/$25 pricing, about 3x the next-best model on ARC-AGI 3, with an effort toggle and a 2.5x-speed Fast mode.
2. GPT-5.6 (Sol / Terra / Luna)
New Frontier FamilyOpenAI's three-tier lineup across ChatGPT, Codex, and the API, from a $5/$30 flagship down to a $1/$6 speedster, with an 80 on the Coding Agent Index, a SOTA 92.2% on BrowseComp, and programmatic tool calling in the Responses API.
3. Kimi K3 (Open Weights)
Open Flagship-Scale ModelMoonshot AI's ~2.8-trillion-parameter open-weight model on Hugging Face with MXFP4 quantization, shipped around July 27 under a revenue-tiered license — one of the largest openly released models to date.
4. Ollama v0.32
Local Agent RuntimeOllama's July release turns local models into tool-using agents with web search, task delegation, a skills system, and expanded GPU support (CUDA on Windows ARM64, NVIDIA B200) — all running on your own hardware.
5. Gemini 3.6 Flash
New Default WorkhorseGoogle's refreshed Flash model at $1.50/$7.50 per million tokens, cutting output tokens ~17% while improving coding — shipped alongside a cheaper Flash-Lite and a security-tuned Flash Cyber.
Discussion