AI News Today: GPT-5.6 & Claude Sonnet 5 Lead a $188B AI Week – July 19, 2026

⚡ Today's AI Briefing (TL;DR)

GPT-5.6 in three tiers: OpenAI's new frontier family — Sol (flagship), Terra (balanced), and Luna (cheapest) — leads with Sol at 80 on the Artificial Analysis Coding Agent Index and 53.6 on Agents' Last Exam, 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.
Open weights arrive from the US: Thinking Machines (Mira Murati's lab) ships Inkling, a 975B-parameter open MoE with near-frontier scores, free to download on Hugging Face and served day-0 by vLLM.
The money is staggering: Databricks raises at a $188B valuation, Fireworks lands a $1.5B round for cheap inference, and SK Hynix completes a record $26.5B Nasdaq IPO — the biggest-ever US listing by a foreign company.

AI News (Top Updates)

1. OpenAI ships the GPT-5.6 family — Sol, Terra, and Luna

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). Sol scores 53.6 on Agents' Last Exam and 80 on the Artificial Analysis Coding Agent Index, while pricing runs from Luna at $1 in / $6 out to Sol at $5 in / $30 out per 1M tokens. OpenAI pitches it as more intelligence per token and stronger performance per dollar.

Why it matters: You can match the model tier to the job — cheap Luna for bulk work, Sol for the hardest coding and agent tasks.

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 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.

Why it matters: The default model most developers reach for just got far more capable without getting more expensive.

3. Thinking Machines opens Inkling, a frontier-class model with full weights

On July 15, Mira Murati's Thinking Machines Lab published Inkling on Hugging Face — a decoder-only multimodal Mixture-of-Experts with ~975B total parameters (41B active), a 1M-token context window, and reasoning across text, image, audio, and video. It ships under an open license with original and NVFP4 checkpoints, and posts GPQA Diamond 87.2%, AIME 2026 97.1%, and SWE-bench Verified 77.6%.

Why it matters: A genuinely open, US-origin frontier model you can self-host and fine-tune — and vLLM already serves it on day zero.

4. Databricks raises a strategic round at a $188B valuation

On July 16, Databricks announced a new strategic funding round led by existing investor Coatue at a $188 billion valuation, earmarked for its Unity AI Gateway, Genie, and Lakebase products plus potential AI acquisitions. It cements Databricks as one of the most valuable private AI-data companies in the world.

Why it matters: Mega-round appetite for AI-data infrastructure is still climbing — and it funds the platforms many teams build on.

5. SK Hynix completes a record $26.5B Nasdaq IPO

On July 10, AI-memory leader SK Hynix listed on Nasdaq, raising roughly $26.5 billion — the largest US IPO ever by a foreign company — and rose about 13% on debut after a ~650% run over the prior year. SK Hynix is the dominant supplier of high-bandwidth memory (HBM) for AI accelerators.

Why it matters: The AI supply chain, not just the model labs, is now a headline market story — and HBM is the bottleneck everyone is watching.

6. xAI ships Grok 4.5 and open-sources Grok Build

In mid-July, xAI released Grok 4.5, a flagship built for coding, agentic tasks, and knowledge work at $2 in / $6 out per 1M tokens, claiming ~4.2x fewer tokens per task (SWE-Bench Pro 64.7%, Terminal-Bench 2.1 83.3%) and shipping into Grok Build, Cursor, and the API. On July 15, xAI also open-sourced Grok Build — its coding agent and terminal UI, including the full agent loop and its skills/plugins/MCP/subagents extension system.

Why it matters: Near-frontier coding keeps getting cheaper, and the agent harness itself is now open for you to run against local inference.

7. Fireworks AI raises $1.5B as the inference-infrastructure boom accelerates

On July 16, AI inference-infrastructure startup Fireworks (Nvidia-backed) closed a $1.505 billion Series D at a $17.5 billion post-money valuation — one of 2026's largest AI-infrastructure rounds. The same week, Together AI raised $800M at $8.3B (Aramco Ventures-led) for the open-model serving layer.

Why it matters: Capital is flooding into the layer that makes running models cheaper and faster — the pipes under your AI apps.

8. NVIDIA and Japan launch the 'world's first national AI infrastructure' on Vera Rubin

On July 16, Japan's government and industrial leaders, with NVIDIA and Noetra, unveiled a nation-scale AI factory for physical AI built on Vera Rubin NVL72 racks (DSX platform): 13,750 Vera CPUs and 27,500 Rubin GPUs at 140 MW, with Spectrum-X Ethernet and BlueField DPUs. It is the first sovereign, nation-scale deployment of NVIDIA's next-gen Rubin generation.

Why it matters: AI is becoming national infrastructure, and NVIDIA's Rubin generation is the backbone countries are building on.

9. The EU gives final green light to simplify and delay the AI Act

On June 29, the EU Council approved an Omnibus package that streamlines the AI Act and pushes key high-risk obligations from August 2, 2026 to December 2, 2027 (embedded systems to August 2028), while adding new bans on AI-generated sexual deepfakes and CSAM. It is a material softening and rescheduling of the world's flagship AI law.

Why it matters: Near-term compliance pressure eases for AI firms operating in Europe — but the rules are delayed, not gone.

10. Major publishers sue Google over Gemini training data

On July 14, Hachette, Cengage, and Elsevier, together with author Scott Turow, sued Google in the Southern District of New York, alleging it used copyrighted books (via Google Books/Play) to train Gemini without permission and altered copyright-management information. The suit extends the defining AI-copyright fight from OpenAI and Anthropic to Google.

Why it matters: The legal question of whether training on copyrighted books is fair use now squarely includes Google's Gemini.

11. Google pushes cheap, agentic, multimodal Gemini across the stack

Google's recent wave includes Nano Banana 2 Lite (gemini-3.1-flash-lite-image), its fastest image model at ~4s and $0.034 per 1K-resolution image, and Gemini Omni Flash for video at $0.10 per second. It also built 'computer use' directly into Gemini 3.5 Flash for browser, mobile, and desktop automation, and shipped the encoder-free Gemma 4 12B open model that runs in ~16GB of memory.

Why it matters: Google is pushing capable multimodal agents both into cheap developer APIs and onto local laptops.

12. Meta expands the Muse family with Spark 1.1, Muse Image, and Muse Video

On July 9, Meta's Superintelligence Labs updated Muse Spark 1.1, a multimodal agentic model with a 1M-token context and gains in tool use, coding, and multi-agent orchestration, now exposed via a new Meta Model API preview. Two days earlier it launched Muse Image (ranked #2 on Arena for text-to-image and editing) and an early-preview Muse Video with native audio. Notably, there were no new Llama-branded model releases in the window.

Why it matters: Meta is doubling down on agentic and creative AI under the Muse brand, even as its open Llama line stays quiet.

13. Mistral pushes into robotics, formal math, and documents

Europe's flagship lab shipped three specialized models: Robostral Navigate (July 8), an 8B embodied navigation model using a single RGB camera that hits 76.6% on R2R-CE; Leanstral 1.5 (July 2), an open-weight Lean 4 proof engine that scores 100% on miniF2F and found 5 unknown bugs across 57 repos; and Mistral OCR 4 (June 23), a 170-language document model at 85.20 on OlmOCRBench.

Why it matters: Mistral is chasing embodied AI, automated math verification, and enterprise document intelligence — not just chat.

14. Agent plumbing matures: MCP goes stateless and open blueprints ship

The Model Context Protocol team released beta SDKs for the 2026-07-28 spec release candidate that make MCP stateless (dropping the initialize handshake) so servers scale horizontally, plus authorization hardening. In parallel, NVIDIA and LangChain shipped the open NemoClaw Deep Agents blueprint, and vLLM added day-0 serving for Inkling and a TileRT integration for latency-critical decode.

Why it matters: The infrastructure for building and scaling reliable agents is getting dramatically less painful.

15. Anthropic and Blackstone bet AI's next trillion is implementation, not models

On July 15, Anthropic and Blackstone (with Hellman & Friedman and Goldman Sachs) launched 'Ode,' a $1.5 billion enterprise AI-implementation firm built around the acquisition of engineering-services startup Fractional AI as a 'Claude-first' foundation. Meanwhile Nebius agreed to sell up to $1B in AI compute to open-source startup Reflection.

Why it matters: The next big AI value pool may be deploying AI inside enterprises — a services layer on top of the models.

Top 5 New / Popular AI Products

1. GPT-5.6 Sol

New Flagship Model

OpenAI's flagship (July 9), scoring 80 on the Artificial Analysis Coding Agent Index and 53.6 on Agents' Last Exam, priced at $5 in / $30 out per 1M tokens across ChatGPT, Codex, and the API.

Why it's trending: Top-tier coding and agent performance with better performance per dollar.

2. Claude Sonnet 5

Agentic Mid-Tier Model

Anthropic'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.

Why it's trending: Frontier-adjacent agent quality at a mid-tier price you can actually ship on.

3. Inkling (Thinking Machines)

Open Frontier Model

A ~975B-parameter multimodal MoE (41B active) with a 1M-token context, released with open weights on Hugging Face (July 15) and served day-0 by vLLM.

Why it's trending: A truly open, US-origin frontier model you can self-host and fine-tune.

4. Grok 4.5

Coding Agent Model

xAI's flagship coding model (mid-July) at $2 in / $6 out per 1M tokens, ~4.2x more token-efficient, shipping into Cursor — with its Grok Build agent now open-sourced.

Why it's trending: Cheap, fast, near-frontier coding help right inside the editor.

5. Nano Banana 2 Lite (Google)

Fast Image Generation

Google's fastest, cheapest image model (June 30) at ~4 seconds and $0.034 per 1K-resolution image, available in AI Studio, the Gemini API, and consumer surfaces.

Why it's trending: Pro-grade image generation at roughly three cents an image.