- GPT-5.6 in three tiers + ChatGPT Work: OpenAI's new frontier family — Sol (flagship), Terra (balanced), and Luna (cheapest) — adds an ultra mode that runs four agents in parallel, priced from $1/$6 to $5/$30 per 1M tokens.
- Claude Sonnet 5 goes agentic: Anthropic's most agentic Sonnet plans, uses browsers and terminals, and runs autonomously — approaching Opus 4.8 quality at $2/$10 intro pricing per 1M tokens.
- Gemini 3.5 Pro slips again: Google's next flagship missed its July 17 deadline after a full architectural rebuild, leaving OpenAI, Anthropic, and xAI to ship uncontested.
- Agents get real plumbing: The 2026-07-28 MCP spec goes stateless so agent servers scale like normal web apps, LangGraph hits 1.0, and DeepSeek V4 brings a 1M-token open-weight context.
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), Terra (balanced), and Luna (cost-efficient), alongside ChatGPT Work for building documents, decks, and websites. A new max reasoning mode and an ultra setting that orchestrates four agents in parallel push ChatGPT toward an agentic work surface. Sol posts 90.4% on BrowseComp (92.2% with Ultra); pricing runs from Luna at $1 in / $6 out to Sol at $5 in / $30 out per 1M tokens.
2. Anthropic releases Claude Sonnet 5, its most agentic Sonnet yet
On June 30, Anthropic launched Claude Sonnet 5 (API id claude-sonnet-5), a mid-tier model that plans, uses tools like browsers and terminals, and runs autonomously — approaching Claude Opus 4.8 quality at a much lower price. It scores 78.5% on OSWorld-Verified (computer use) and 46.8% on Humanity's Last Exam with tools. Introductory pricing is $2 in / $10 out per 1M tokens through August 31, then $3 / $15.
3. Google's Gemini 3.5 Pro misses its deadline again
After a full architectural rebuild and a scrapped base model, Google's next flagship reportedly slipped past its July 17 target and remained unshipped as of mid-July, with the company weighing a stopgap release. Bloomberg reports the model fell short of internal goals — a rare public stumble for the frontier leader while OpenAI, Anthropic, and xAI all shipped.
4. xAI ships Grok 4.5, adds Automations and open-sources Grok Build
On July 16, xAI (now SpaceXAI) released Grok 4.5, billed as its smartest model for coding, agentic tasks, and knowledge work, plus a new Automations feature for scheduled, long-running tasks. On July 15 it open-sourced Grok Build — its coding agent and terminal UI, including the full agent loop and its skills, plugins, MCP, and subagents extension system.
5. GPT-5.6 becomes Microsoft 365 Copilot's preferred model
On July 9, Microsoft made GPT-5.6 the default engine in Microsoft 365 Copilot, and it is merging its consumer and enterprise Copilot apps into a single application while launching paid, autonomous AutoPilot agents. Copilot also gained multimodal Vision, email-drafting refinement, and Excel, SharePoint, and Search upgrades this month.
6. MCP's biggest overhaul lands as a release candidate
The Model Context Protocol team shipped beta SDKs for the 2026-07-28 spec release candidate — the largest revision since launch. It goes stateless (dropping the initialize handshake) so servers scale behind standard load balancers, and adds Multi Round-Trip Requests, server-rendered MCP Apps UIs, a long-running Tasks extension, and OAuth-aligned authorization.
7. DeepSeek pushes V4 toward a formal release with 1M-token context
Following its April preview of V4-Pro (1.6T total / 49B active parameters) and V4-Flash (284B / 13B active), DeepSeek is retiring the legacy deepseek-chat and deepseek-reasoner endpoints on July 24 and reportedly moving to peak-hour API pricing that doubles during Beijing work blocks. All V4 models offer a 1M-token context via DeepSeek Sparse Attention and ship open weights on Hugging Face.
8. NVIDIA opens robotics and edge AI with Jetson Thor
On July 15, NVIDIA introduced new Jetson Thor compute modules that run foundation models on-device for robots and edge deployments, alongside its expanding Nemotron open-model family, which Japanese enterprises are using to build industry-specialized AI. NVIDIA also touted performance-per-watt as the metric that now defines AI infrastructure efficiency.
9. Meta expands the Muse family with Spark 1.1, Muse Image, and Muse Video
On July 9, Meta updated Muse Spark 1.1, a multimodal agentic model. Two days earlier it launched Muse Image, which follows instructions faithfully and composes from multiple references, and an early-preview Muse Video with native audio. Notably there were no new Llama-branded model releases in the window, and reports suggest Meta's next Llama flagship is delayed amid a reorganization.
10. Google renames NotebookLM to Gemini Notebook
On July 16, Google rebranded its research tool to Gemini Notebook, tying it to the Gemini family with a new gradient logo. Each notebook now gets a secure cloud computer for executing code and grounded data analysis, and the tool is expanding into the Gemini app and Search's AI Mode. The Gemini 3.5 and Antigravity upgrade, once limited to AI Ultra, rolls out to AI Pro users over the coming weeks.
11. Mistral pushes into robotics, formal math, and documents
Europe's flagship lab shipped several specialized models: Robostral Navigate (July 8), an embodied navigation model using a single camera; Leanstral 1.5 (July 2), an open-weight Lean 4 proof engine; and Mistral OCR 4 (June 23), a state-of-the-art document-intelligence model. It also added a system of record for prompts and skills in Studio with versioning and traceability.
12. Open reasoning models close the gap on the Hugging Face Hub
Hugging Face's ecosystem keeps compounding: NVIDIA and Hugging Face released OpenReasoning-Nemotron, a family of distilled, state-of-the-art reasoning models, and the fully open SmolLM3-3B multilingual long-context reasoner punches far above its size. The broader open field — DeepSeek V4, Qwen 3.7, GLM-5.2, and Kimi K2.6 — continues to narrow the distance to closed frontier models.
13. Agent tooling matures: LangGraph 1.0 and Ollama move forward
LangChain's agent-orchestration framework LangGraph reached its 1.0 GA milestone, giving developers a stable foundation for durable, stateful agents. The local runtime Ollama shipped a new release improving support and performance for the latest open models, and NVIDIA's Nemotron topped the LangChain Deep Agents benchmark harness on July 8.
14. Anthropic tops OpenAI on valuation as AI races into classrooms
Anthropic raised roughly $65B at a ~$965B valuation in late May, briefly becoming the most valuable AI startup and nearing the $1T mark. It also launched Claude for Teachers (July 14) and committed $10M to Canadian AI research, while the U.S. eased export controls affecting its Claude Fable 5 and Mythos 5 models.
Top 5 New / Popular AI Products
1. GPT-5.6 Sol
New Flagship ModelOpenAI's flagship (July 9), scoring 90.4% on BrowseComp (92.2% with Ultra) with a new ultra multi-agent mode, priced at $5 in / $30 out per 1M tokens across ChatGPT, Codex, and the API.
2. Claude Sonnet 5
Agentic Mid-Tier ModelAnthropic's most agentic Sonnet (June 30) — plans, uses browsers and terminals, and runs autonomously near Opus 4.8 quality at $2 in / $10 out per 1M intro pricing.
3. Grok 4.5
Coding Agent ModelxAI's flagship built for coding, agentic tasks, and knowledge work (July 16), shipping with a new Automations feature and the now open-sourced Grok Build agent.
4. DeepSeek V4
Open-Weight Frontier ModelDeepSeek's V4-Pro (1.6T total / 49B active) and V4-Flash offer a 1M-token context via Sparse Attention with open weights on Hugging Face; legacy endpoints retire July 24.
5. Gemini Notebook
Rebranded Research ToolGoogle's renamed NotebookLM (July 16) now gives each notebook a secure cloud computer for code execution and grounded data analysis, expanding into the Gemini app and Search AI Mode.
Discussion