- GPT-5.6 in three tiers + ChatGPT Work: OpenAI's frontier family — Sol, Terra, and Luna — adds an ultra mode that runs agents in parallel plus Programmatic Tool Calling, priced from $1/$6 to $5/$30 per 1M tokens.
- Grok 4.5 goes 'Opus-class': xAI's new coding model hits 64.7% on SWE-Bench Pro at $2/$6 per 1M tokens and ships into Cursor, with Grok Build open-sourced a day earlier.
- Anthropic's $1.5B copyright settlement clears: a federal judge approved the largest known U.S. copyright settlement over books used to train Claude, as publishers separately sued Google over Gemini.
- Open weights and new rules: Meta pivots to closed 'Muse' models while Inkling, Kimi K3, and DeepSeek V4 push open trillion-scale AI — and the EU AI Act reaches full applicability on August 2.
AI News (Top Updates)
1. OpenAI ships the GPT-5.6 family and ChatGPT Work
On July 9, OpenAI released GPT-5.6 in three tiers across ChatGPT, Codex, and the API — Sol (flagship, with an ultra mode that runs agents in parallel), Terra (balanced), and Luna (budget) — alongside ChatGPT Work, an enterprise agent that builds documents, spreadsheets, and presentations. The Responses API also gained Programmatic Tool Calling, which lets the model write JavaScript in an isolated runtime and cut tokens 38–63.5% for named customers. Sol scored 80 on the Artificial Analysis Coding Agent Index; pricing runs from Luna at $1 in / $6 out to Sol at $5 in / $30 out per 1M tokens.
2. A federal judge approves Anthropic's $1.5B author copyright settlement
On July 21, a San Francisco federal judge granted final approval of Anthropic's settlement with authors whose books were used to train Claude — what the court called the largest known settlement of a U.S. copyright case. Over 91% of covered rightsholders filed claims, and attorneys were awarded more than $101M in fees. It follows the earlier ruling that model training was fair use but that storing pirated books was not.
3. xAI ships Grok 4.5 and open-sources Grok Build
On July 16, xAI released Grok 4.5, billed as an 'Opus-class' model for coding, agentic tasks, and knowledge work — scoring 64.7% on SWE-Bench Pro and 83.3% on Terminal-Bench 2.1 at $2 in / $6 out per 1M tokens, and available in Cursor on all plans. A day earlier it open-sourced Grok Build, its agent and development tooling, and it also shipped 21 new flagship voices.
4. Meta pivots from open Llama to closed 'Muse' models
Meta Superintelligence Labs launched Muse Spark 1.1 (July 9), a multimodal agentic model with a 1M-token context, behind a new paid Meta Model API — effectively its first closed-weight frontier model. Two days earlier it shipped Muse Image (ranked No. 2 on Arena for text-to-image and editing) and previewed Muse Video with native audio. No new open Llama model shipped in the window.
5. Japan turns on the world's first national AI infrastructure on NVIDIA
On July 16, Japan's government and industrial partners deployed a sovereign 'physical AI' build with NVIDIA using 13,750 Vera CPUs and 27,500 next-generation Rubin GPUs. NVIDIA also introduced Jetson Thor edge modules (July 15) for on-device robotics and a capital-partner model (July 1) that funds multi-tenant 'AI factories' via revenue sharing.
6. MCP's biggest overhaul lands as a release candidate
The Model Context Protocol team shipped a release candidate for the 2026-07-28 spec, its largest revision yet. It goes stateless — dropping the initialize handshake and session management so agent servers scale behind ordinary load balancers — and adds official MCP Apps (sandboxed HTML UIs), a redesigned Tasks extension, and tightened OAuth authorization. Roots, Sampling, and Logging get a 12-month deprecation window.
7. Open-weight, trillion-scale models keep escalating
Thinking Machines released Inkling on Hugging Face (July 15) — an open ~1T-parameter (41B active) natively multimodal MoE with a 1M-token context that scores 87.2% on GPQA Diamond. Moonshot previewed Kimi K3 (a 2.8T MoE, weights due July 27) and Alibaba previewed Qwen3.8-Max (claimed 2.4T parameters) at Shanghai's World AI Conference, while DeepSeek's open-source V4 (1M-token context) reached its formal release.
8. The EU AI Act reaches full applicability on August 2
In under two weeks, the EU AI Act becomes fully applicable, activating chatbot-transparency duties (disclosing when users interact with AI) and general-purpose-AI obligations around training-data transparency, copyright compliance, and systemic-risk assessment. High-risk-system rules remain staggered into 2027 and 2028.
9. Microsoft leans on its own MAI models and opens Foundry to Claude
Microsoft's first-party MAI family — seven models spanning reasoning, coding, image, transcription, and voice — reduces its reliance on outside labs inside Copilot, with MAI-Code-1-Flash (5B active) wired into GitHub Copilot and VS Code. Anthropic's Claude also reached general availability in Microsoft Foundry (July 7), which can now publish governed 'autopilot' agents into Microsoft 365 Copilot and Teams.
10. Anthropic launches Claude for Teachers
On July 14, Anthropic released a free tool that maps to academic standards from all 50 U.S. states, builds lesson plans, and analyzes assessment data — built with the American Federation of Teachers and excluded from model training. A classroom pilot launches next school year in the Detroit Public Schools Community District.
11. Google DeepMind lays out a 'bioresilience' plan and adds Computer Use to Gemini
On July 16, DeepMind published a framework for sharing models with trusted partners to prevent misuse, detect outbreaks, and speed response — including adapting SynthID watermarking to DNA so synthesis providers can screen risky AI-generated sequences. Separately, Gemini 3.5 Flash gained agentic 'computer use,' letting it see, reason, and act across desktop, mobile, and browser.
12. Mistral pushes into robotics, formal math, and documents
Europe's flagship lab shipped several specialized models: Robostral Navigate (July 8), its first embodied navigation model; Leanstral 1.5 (API June 30), an open Lean 4 proof engine; and Mistral OCR 4 (June 23), a document-intelligence model. It also added a system of record for prompts and skills in Studio, with versioning and traceability.
13. Open agent tooling matures: vLLM, Ollama, and Hugging Face's LeRobot
vLLM cut v0.25.0 (July 11), making Model Runner V2 the default for dense models and retiring legacy PagedAttention, with the Transformers backend now 'as fast as native vLLM.' Ollama v0.32.1 (July 16) improved Gemma 4 tool calling and fixed an MLX memory leak, and Hugging Face's LeRobot v0.6.0 added world-model policies, five new VLAs, and cloud training via HF Jobs.
14. The money and law around AI both move: Databricks, Google, and compute
Databricks announced a Coatue-led strategic round valuing it at $188B (July 16–17). The same week, Hachette, Cengage, Elsevier, and author Scott Turow sued Google, alleging it copied books — including from pirate sources — to train Gemini (July 15), and CNBC reported Anthropic is in early talks to buy compute capacity from Meta. For context, Anthropic earlier closed a ~$65B round at a ~$965B valuation.
Top 5 New / Popular AI Products
1. GPT-5.6 Sol
New Flagship ModelOpenAI's flagship (July 9), scoring 80 on the Artificial Analysis Coding Agent Index with an ultra multi-agent mode, priced at $5 in / $30 out per 1M tokens across ChatGPT, Codex, and the API.
2. Grok 4.5
Coding Agent ModelxAI's 'Opus-class' model for coding and agentic work (July 16) — 64.7% on SWE-Bench Pro at $2 in / $6 out per 1M tokens, with the open-sourced Grok Build agent alongside it.
3. Muse Spark 1.1
Agentic Multimodal ModelMeta Superintelligence Labs' first paid, closed-weight agentic model (July 9), with a 1M-token context behind the new Meta Model API and in Meta AI's Thinking mode.
4. DeepSeek V4
Open-Weight Frontier ModelDeepSeek's V4-Pro (1.6T total / 49B active) and V4-Flash (284B / 13B active) offer a 1M-token context via Sparse Attention with open weights on Hugging Face, plus new time-of-day API pricing.
5. Inkling
Open 1T Multimodal ModelThinking Machines' open ~1T-parameter (41B active) natively multimodal MoE on Hugging Face (July 15), with a 1M-token context and 87.2% on GPQA Diamond, runnable via vLLM, SGLang, and llama.cpp.
Discussion