- Claude Opus 5 headlines the week: Anthropic's new flagship (
claude-opus-5) delivers frontier reasoning and coding, holds pricing flat at $5/$25 per million tokens, claims roughly 3x the next-best model on ARC-AGI 3, and adds a 2.5x-speed Fast mode. - The GPT-5.6 vs Gemini Flash reset: OpenAI's GPT-5.6 family (Sol/Terra/Luna, from $5/$30 down to $1/$6) posts big agent gains, while Google's Gemini 3.6 Flash becomes the new default workhorse using ~17% fewer output tokens at $1.50/$7.50.
- Open weights surge: Thinking Machines' ~1T-parameter multimodal Inkling and Moonshot's Kimi K3 both ship openly on Hugging Face, pushing frontier-scale models into download-and-run territory alongside DeepSeek V4's 1M-context MoE.
- NVIDIA stands up a national AI factory: Japan switches on a sovereign supercomputer with 13,750 Vera CPUs and 27,500 Rubin GPUs as Vera Rubin NVL72 enters production across the major clouds.
AI News (Top Updates)
1. Anthropic launches Claude Opus 5, its new flagship
Released July 24, Claude Opus 5 (claude-opus-5) brings major gains in reasoning, self-verification, coding, knowledge work, and scientific research, along with improved alignment. 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 new Fast mode running about 2.5x faster for double the base price.
2. 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 multiple agents in parallel on the hardest tasks. Sol scores 80 on the Artificial Analysis Coding Agent Index (up from GPT-5.5's 76.4) while using under half the output tokens, and posts a state-of-the-art 92.2% on BrowseComp with up to 1M-token context. Pricing per million tokens runs Sol $5/$30, Terra $2.50/$15, and Luna $1/$6.
3. Google refreshes the Gemini Flash tier
On July 21 Google shipped Gemini 3.6 Flash — the new default workhorse, cutting output tokens about 17% while improving coding at $1.50/$7.50 per million — alongside Gemini 3.5 Flash-Lite ($0.30/$2.50, running 350 tokens per second and scoring 54.2% on SWE-Bench Pro) and Gemini 3.5 Flash Cyber, a cybersecurity-tuned model paired with CodeMender. Google also committed $40M in AI tokens, Cloud credits, and Gemini for Government seats to the U.S. Department of Energy's Genesis Mission on July 22.
4. xAI ships Grok 4.5 for coding and agents
Released July 16, Grok 4.5 is built for coding, agentic tasks, and knowledge work, with strong results on hard languages like Rust and C/C++ and end-to-end app development. It posts 83.3% on Terminal Bench 2.1, 64.7% on SWE-Bench Pro, and 62.0% on DeepSWE 1.0, with API pricing of $2 per million input and $6 per million output tokens and claimed ~4.2x token efficiency. xAI also launched Automations in Grok for scheduling recurring tasks, and open-sourced its Grok Build coding environment.
5. The Model Context Protocol goes stateless
The MCP 2026-07-28 specification — a release candidate since May 21 and scheduled to finalize 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.
6. NVIDIA stands up the first national-scale AI factory
On July 16 NVIDIA, the Japanese government, and industry partners launched what NVIDIA calls the world's first national AI infrastructure, built on its next-generation Vera Rubin platform and deployed via Noetra Corp — 13,750 Vera CPUs and 27,500 Rubin GPUs. Days later, Vera Rubin NVL72 entered production ramp across CoreWeave, Google Cloud, Microsoft Azure, and Oracle Cloud Infrastructure, and NVIDIA introduced Spectrum-6 networking built for AI factories spanning hundreds of thousands of processors.
7. 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 and quantized NVFP4 formats with day-zero support in transformers, vLLM, SGLang, and llama.cpp.
8. Moonshot's Kimi K3 pushes open weights toward flagship scale
Moonshot AI began releasing Kimi K3 open weights on its official Hugging Face org, with the model page live as an upcoming release and weights dropping around July 27. Community overviews place it in the ~2.8-trillion-parameter class with MXFP4 quantization, making it one of the largest openly released models to date and intensifying the open-weight race out of China alongside DeepSeek and Alibaba's Qwen line.
9. DeepSeek V4 lands as a low-cost, long-context open option
DeepSeek's V4 preview ships in two variants — V4-Pro (1.6T total / 49B active) and V4-Flash (284B total / 13B active) — and makes a 1-million-token context window the default across DeepSeek's services. It uses token-wise compression plus DeepSeek Sparse Attention for cheap long-context inference and is compatible with both OpenAI Chat Completions and Anthropic-style APIs, positioning it as a frontier-class option at a fraction of typical cost.
10. Meta ships Muse Spark 1.1 and Muse Image
Meta Superintelligence Labs released Muse Spark 1.1 on July 9 via the new Meta Model API — a multimodal reasoning model for agentic tasks (tool and computer use, advanced coding) with a 1M-token context window, parallel tool calling, and built-in search with citations, available in "Thinking" mode in the Meta AI app. Days earlier, Meta launched Muse Image, its first in-house image generator, inside Meta AI, with reasoning-based generation, precise editing, multi-photo blending, and legible in-image text.
11. Microsoft brings Claude to Foundry and expands Azure compute with AMD
Anthropic's Claude models are now generally available on Microsoft Foundry via the Messages API — with prompt caching, extended thinking, and tool streaming — and can serve as the reasoning core for Foundry Agent Service, while Foundry agents can publish directly to Teams and Microsoft 365 Copilot without rebuilds. On July 20, Microsoft announced it will deploy AMD's Instinct MI455X GPUs (Helios rack-scale) and 6th-gen EPYC "Venice" CPUs across Azure AI data centers, diversifying its compute beyond NVIDIA.
12. Mistral pushes into physical AI and enterprise governance
On July 8 Mistral released Robostral Navigate, an 8B vision-language model that lets robots navigate from natural-language instructions using a single RGB camera — no LiDAR or depth sensors — hitting 76.6% success on R2R-CE unseen validation after training on about 400,000 simulated trajectories across 6,000 scenes. Mistral also positioned Mistral Studio as a versioned "system of record" for prompts and skills, and shipped Leanstral 1.5 for formal proof and math reasoning.
13. Business & policy: mega-rounds, AMD's rackscale push, and U.S. AI procurement
Enterprise-AI infrastructure company Fireworks AI raised a $1.5B Series D at a $17.5B valuation (Atreides, Index Ventures, TCV), and AI drug-discovery firm Chai Discovery raised a $400M Series C at $3.8B. On hardware, AMD's "Advancing AI 2026" unveiled the Instinct MI400 series, the Helios rackscale system (72 MI455X GPUs plus 18 EPYC CPUs), and 6th-gen EPYC "Venice" CPUs with up to 256 cores — its most direct challenge yet to NVIDIA's Rubin. In Washington, contractors and experts said the U.S. GSA's revised draft AI procurement rule still falls short on transparency and testing obligations.
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 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 and a SOTA 92.2% on BrowseComp.
3. Grok 4.5
New Coding ModelxAI's coding-and-agents flagship, scoring 83.3% on Terminal Bench 2.1 and 64.7% on SWE-Bench Pro at $2/$6 per million tokens, available in the console, Cursor, and Grok Build.
4. Thinking Machines Inkling
Open Multimodal ModelA ~1-trillion-parameter (≈41B active) open MoE multimodal model on Hugging Face, handling text/image/audio with 1M-token context and day-zero support in transformers, vLLM, SGLang, and llama.cpp.
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 350-tokens/sec Flash-Lite and a security-tuned Flash Cyber.
Discussion