- Three flagships, one window: OpenAI's GPT-5.6 family (Sol, Terra, Luna) plus agentic ChatGPT Work, xAI's Grok 4.5, and Anthropic's Claude Sonnet 5 all landed within days of each other.
- Meta opens up: Muse Spark 1.1 arrives with the new OpenAI-compatible Meta Model API and a 1M-token context — the first real developer access to Meta's frontier models.
- Enterprise + compute: Microsoft made Anthropic's Claude models GA on Azure Foundry, and Nvidia + Japan announced the world's first national AI infrastructure on the Vera Rubin platform.
- Agents get a standard: The Model Context Protocol's 2026-07-28 release candidate goes stateless, making agent servers scale like ordinary web services.
AI News (Top Updates)
1. OpenAI ships the GPT-5.6 family (Sol, Terra, Luna) and reframes ChatGPT as a work agent
On July 9, OpenAI released GPT-5.6 in three tiers — Sol (flagship), Terra (balanced), and Luna (cost-efficient) — adding an ultra mode that coordinates multiple agents in parallel and "Programmatic Tool Calling" for in-memory tool coordination. Pricing runs from Luna at $1 in / $6 out to Sol at $5 in / $30 out per 1M tokens. Alongside it, "ChatGPT Work" turns ChatGPT into an agent that acts across apps and files, runs for hours, and handles Scheduled Tasks and desktop Computer Use.
2. Anthropic launches Claude Sonnet 5 — "the most agentic Sonnet yet"
Announced June 30, Claude Sonnet 5 posts big gains over Sonnet 4.6 in reasoning, tool use, and coding — approaching Opus-class capability at a mid-tier price, with lower hallucination and sycophancy. Introductory pricing is $2 in / $10 out per 1M tokens through August 31 (then $3 / $15), under the model ID claude-sonnet-5.
3. xAI releases Grok 4.5, its "Opus-class" coding model
On July 8, xAI shipped Grok 4.5, "built for coding, agentic tasks, and knowledge work." It scores 64.7% on SWE Bench Pro and 83.3% on Terminal Bench 2.1, runs at roughly 80 tokens/sec with claimed 4.2× token efficiency, and is priced at $2 in / $6 out per 1M tokens — available in Grok Build, Cursor, and the xAI API.
4. Meta opens the Muse family with Muse Spark 1.1 and the Meta Model API
On July 9, Meta Superintelligence Labs released Muse Spark 1.1, a multimodal reasoning/agentic model, alongside the new Meta Model API — an OpenAI-compatible endpoint with structured output and parallel tool calling. It offers a 1M-token context window, orchestrates parallel subagents, and powers a "Thinking" mode in the Meta AI app.
5. GPT-5.6 becomes the preferred model inside Microsoft 365 Copilot
The same day it launched (July 9), Microsoft made GPT-5.6 the default reasoning model powering Microsoft 365 Copilot — putting the newest frontier model in front of hundreds of millions of enterprise seats immediately, deepening the OpenAI–Microsoft distribution channel on launch day.
6. Anthropic's Claude models hit general availability on Azure / Microsoft Foundry
Microsoft made Claude generally available via the Messages API with prompt caching, extended thinking, and tool streaming, and Foundry Agent Service can now use Claude as its reasoning core. It ships with global and US data zones, a zero-data-retention option, and consolidated billing via Claude Consumption Units.
7. Nvidia and Japan launch "the world's first national AI infrastructure for physical AI"
On July 16, Nvidia and Noetra, backed by Japan's government, unveiled a national AI factory built on the next-generation Vera Rubin platform: 13,750 Vera CPUs and 27,500 Rubin GPUs, 140 MW, on the DSX architecture with Vera Rubin NVL72 racks and Spectrum-X Ethernet. Nvidia also debuted lower-cost Jetson Thor T2000/T3000 edge modules for mainstream robotics.
8. xAI open-sources Grok Build, its coding agent and terminal UI
On July 15, xAI released the source for Grok Build — its coding agent and terminal UI, including the agent loop, edit tools, a diff-viewing TUI, and an extension system for skills, plugins, hooks, MCP, and subagents. The code is published at github.com/xai-org/grok-build.
9. The Model Context Protocol goes stateless in its biggest revision yet
MCP published the 2026-07-28 spec release candidate — its largest revision since launch. The protocol becomes stateless (dropping the initialize handshake and session IDs, so any request can hit any server behind a load balancer) and adds a formal Extensions framework, including MCP Apps (server-rendered UIs) and Tasks, plus JSON Schema 2020-12 tool schemas and W3C Trace Context.
10. Mistral pushes into physical AI with Robostral Navigate and ships prompt governance
On July 8, Mistral introduced Robostral Navigate, an 8B open embodied-navigation model that steers wheeled, legged, and flying robots from a single RGB camera (no LiDAR), hitting 76.6% success on unseen validation after training on ~400,000 simulated trajectories. A day later, Mistral AI Studio added Prompt Management — a versioned, ownable "system of record" for prompts and skills.
11. Hugging Face ships LeRobot v0.6.0 and a revamped Kernels project
LeRobot v0.6.0 (July 7) adds imagination-based world-model policies, new vision-language-action models (GR00T N1.7, MolmoAct2), six new sim benchmarks, and FSDP multi-GPU training. The Kernels revamp (July 6) makes custom compute kernels first-class Hub repos with trusted-publisher verification and Sigstore code signing.
12. Google expands the open + multimodal Gemini stack
Google's recent updates include Gemma 4 12B, an open model that runs locally on 16GB with vision and native voice; Nano Banana 2 Lite, its "fastest and most cost-efficient Gemini Image model yet"; and Gemini Omni Flash arriving in the APIs in public preview, enabling custom dynamic video workflows — alongside Gemini 3.5 Flash Computer Use and Live Translate across 70+ languages.
13. Open-weight frontier: DeepSeek-V4-Pro and Qwen3.6 lead the pack
DeepSeek-V4-Pro is a 1.6T-parameter MoE (49B active) with a 1M-token context, released open-weight under the MIT license with hybrid attention and three reasoning modes. Alibaba's Qwen3.6 series (35B-A3B MoE, ~3B active) ships under Apache 2.0 with a focus on agentic coding — and vLLM v0.22.x has matured serving for both with NVFP4 MoE and multi-tier KV-cache offloading.
14. Media gen levels up: Runway Agent Skills, FLUX.2 on-device, ElevenLabs SynthID detection
Runway shipped Agent Skills (one-command ad creation and localization), Seed Audio 1.0 (up to 120s of speech/sound/music from a prompt), and Agent 2.0. Black Forest Labs put FLUX.2 "Klein" on ASUS ProArt laptops via an ASUS/Nvidia partnership. ElevenLabs added SynthID-based detection of its own generated audio, tied to Google DeepMind's watermarking standard.
15. Business & policy: Together AI's $800M, ElevenLabs' ~$22B tender, and a softer EU AI Act
Together AI raised an $800M Series C at $8.3B (led by Aramco Ventures), and ElevenLabs is reportedly in talks for a tender offer near $22B. The EU Council gave final green light to the AI Act "Digital Omnibus," deferring high-risk compliance to Dec 2, 2027 while adding a December 2026 ban on non-consensual "nudifier" deepfakes. A reality check: on the new Long-Horizon-Terminal-Bench, the best of 15 models scored just ~15%.
Top 5 New / Popular AI Products
1. Claude Sonnet 5
New Mid-Tier ModelAnthropic's newest workhorse model for coding, agents, and professional work, launched June 30 at $2/$10 per 1M tokens — approaching Opus-class capability at the tier most production apps actually run on.
2. Meta Model API
OpenAI-CompatibleMeta's new developer endpoint (July 9) exposes Muse Spark 1.1 with structured output, parallel tool calling, and a 1M-token context — all behind an OpenAI-compatible interface.
3. Ollama v0.30.9 (Hermes Desktop)
Local AI Agents
Recent Ollama builds add a native "Hermes Desktop" agent UI (ollama launch hermes-desktop), Cohere2Moe architecture support, and Gemma 4 with QAT — turning the local runner into an agent-capable platform.
4. Grok Build (open source)
Coding AgentxAI open-sourced its coding agent and terminal UI (July 15) — agent loop, edit tools, a diff-viewing TUI, and an extension system for skills, plugins, hooks, MCP, and subagents.
5. Mistral Studio Prompt Management
Prompt System-of-RecordMistral Studio (July 9) gives AI prompts and skills a versioned, owned, and traceable system of record — governance tooling for teams shipping AI in production.
Discussion