Computex Week: AI Just Moved Onto Your Desk — and Into Your Office Stack.

Claude Opus 4.8 lands with Dynamic Workflows. Microsoft turns OpenClaw into an enterprise agent called Scout. NVIDIA + Microsoft put 120B-parameter models on a laptop. MiniMax ships the first open-weights model with all three frontier capabilities. NVIDIA Cosmos 3 opens up physical AI. Plus 15+ tools, agents, and 10 GitHub repos you should know.

Published: June 3, 2026 · Coverage window: May 20 – June 3

Introduction

Two weeks ago, the headline was "frontier coding at 1/10th the cost." This round, the headline is bigger and harder to summarize: the entire physical layer of AI just shifted. Anthropic shipped Opus 4.8 with the ability to orchestrate hundreds of sub-agents over codebase-scale work. Microsoft made OpenClaw — an open-source agent framework — the foundation of its first enterprise "Autopilot" agent across Microsoft 365. NVIDIA and Microsoft jointly announced a Windows-on-Arm platform that runs 120-billion-parameter models on a laptop. And MiniMax open-sourced a model that ties frontier coding, 1-million-token context, and native multimodality into one architecture.

For GCC teams, this round is the moment to rethink where your AI runs. Three years of "everything in the cloud" is being inverted in real time. Local inference, hybrid orchestration, and on-device agents are no longer the future — they're the production stack for late 2026.

01 · Frontier: Claude Opus 4.8 ships with Dynamic Workflows

MAY 28 · Codebase-scale migrations · API: claude-opus-4-8

Anthropic released Claude Opus 4.8 — an upgrade to Opus 4.7 with sharper judgment, more honesty about its own progress, and the ability to work independently for longer than its predecessors. Available through claude.ai, Claude Code, and the Claude API as claude-opus-4-8, at the same price as Opus 4.7.

The headline feature isn't a benchmark — it's a new capability called Dynamic Workflows. The system plans work, runs parallel sub-agents, verifies outputs, and reports back. Anthropic's framing: Claude Code alongside Opus 4.8 can now carry out codebase-scale migrations across hundreds of thousands of lines of code from kickoff to merge, with the existing test suite as its bar.

What changed in this release:

  • Dynamic Workflows — Opus 4.8 plans complex work into verifiable subtasks, runs parallel sub-agents, validates against tests, and only merges when green. Available in research preview for Claude Code Enterprise, Team, and Max users.
  • Effort Control — Token-Burn Slider. In claude.ai and Cowork, you can now set how much effort Claude applies — /effort xhigh for the hardest tasks.
  • Fast Mode — 2.5× faster, at $10/M input and $50/M output — roughly 3× cheaper than the previous fast-mode generation.
  • Messages API Live Edits — The Messages API now accepts live changes to the messages array mid-task without breaking prompt cache.

For Conneqt builds: Dynamic Workflows is the right primitive for Second Brain agentic flows and Hermes-style orchestration. Instead of a single long prompt chain that drifts, you get parallel sub-agents reporting against verifiable criteria. Worth testing on a real client workflow this week.

One more note: Anthropic confirmed Mythos-class models will reach all customers in the coming weeks once cybersecurity safeguards are finalized. The Mythos preview under Project Glasswing remains restricted to a small set of trusted organizations.

02 · Open Weights: MiniMax M3 — first open-weights model to combine three frontier capabilities

JUNE 1 · 1M-token context · 59.0% SWE-BENCH PRO · $0.60/M input

Shanghai-based MiniMax released M3 on June 1 — positioned as the first open-weight model to fuse frontier-level coding, a 1M-token context window, and native multimodality (image + video + desktop computer operation) in a single architecture. API live now; weights committed to Hugging Face within ~10 days.

Benchmark snapshot (vendor-reported):

BenchmarkM3What it measures
SWE-Bench Pro59.0%Above GPT-5.5 (58.6%), approaching Opus 4.7
Terminal-Bench 2.166.0%Shell-centric agentic workflows
SWE-fficiency34.8%Cost-aware engineering tasks
KernelBench Hard28.8%Low-level systems code
MCP Atlas74.2%Model Context Protocol tool use

The architectural story is MiniMax Sparse Attention (MSA). Standard attention is quadratic — context cost grows as the square of sequence length. MSA scales to 1M tokens with reported 9.7× faster prefill and 15.6× faster decoding. Pricing lands at $0.60/M input — roughly 12× cheaper than Opus 4.8 on input.

Caveats worth keeping: Benchmarks are vendor-run, not yet independently verified. Weights had not shipped at time of writing. For teams processing sensitive client data through the API, note that Chinese law (the 2017 National Intelligence Law) gives the Chinese government legal access to data processed by MiniMax's API. Self-hosting the weights once published is the path that resolves this.

03 · Physical AI: NVIDIA Cosmos 3 — the first fully open omnimodel for physical AI

JUNE 1 · Hugging Face · SUPER 32B + NANO 8B

Announced at NVIDIA GTC Taipei alongside Computex, Cosmos 3 is the world's first fully open omnimodel — one architecture that natively understands and generates text, images, video, ambient sound, and physical actions (joint angles, gripper positions, trajectory waypoints). Built on a mixture-of-transformers architecture, trained on 20 trillion multimodal tokens.

What was released:

  • Cosmos 3 Super — 32B reasoner + 32B generator, available now on Hugging Face
  • Cosmos 3 Nano — 8B reasoner + 8B generator, available now
  • Cosmos 3 Edge — coming soon for real-time inference on-device
  • NVIDIA Cosmos Coalition — partners include Agile Robots, Black Forest Labs, Generalist, LTX, Runway, Skild AI, Doosan Robotics, LG, Samsung, Li Auto

Jensen Huang's framing: "The big bang of physical AI is just around the corner." For GCC industrial, logistics, and smart-city deployments (NEOM, KSA Vision 2030, port automation), this is the open foundation that makes regional robotics development viable without renting cloud GPUs by the hour.

04 · Enterprise: Microsoft Scout — the first "Autopilot" agent, built on OpenClaw

JUNE 2 · BUILD 2026 · M365 integration · OpenClaw base

Microsoft announced Scout at Build 2026 — the first of a new agent category Microsoft calls "Autopilots." Always-on, autonomously acting, with its own governed identity, integrated across Teams, Outlook, OneDrive, and SharePoint. Microsoft built Scout on top of the open-source OpenClaw framework — 180,000+ GitHub stars in three months — and committed to contributing enterprise-grade policy controls back upstream.

Scout schedules meetings across time zones, flags stalled decisions, blocks calendar time before deadlines hit, and runs in the background using Work IQ as its context engine.

Capabilities:

  • Distribution — Rolling out now via Microsoft 365 Frontier. Available inside the Copilot app and as desktop apps on Windows and macOS.
  • Surface area — Teams + Outlook + OneDrive + SharePoint. Extends to browser and local resources via MCP servers.
  • Governance — Policy conformance system continuously checks whether the agent is operating within guidelines.
  • Upstream — Microsoft committing enterprise policy controls back to the OpenClaw open-source project.

The strategic read: This is the single most important enterprise AI move of 2026 so far. Microsoft validated open-source agent infrastructure as the path forward. For GCC enterprises already on Microsoft 365 (banks, ministries, telcos, healthcare networks), Scout is the lowest-friction path to deploy production agents inside an already-governed stack.

05 · Hardware: NVIDIA RTX Spark + Surface Laptop Ultra — 120B-parameter models on a laptop

COMPUTEX · MAY 31 · 128GB unified memory · 1 petaflop AI

NVIDIA unveiled the RTX Spark Superchip at Computex Taipei — a Windows-on-Arm platform pairing a 20-core Grace CPU with a Blackwell GPU (6,144 CUDA cores) and up to 128GB of unified LPDDR5X memory with 300 GB/s bandwidth, linked by NVLink-C2C. Run 120-billion-parameter LLMs with up to 1M-token context locally, plus 1 petaflop of FP4 AI compute.

Microsoft's Surface Laptop Ultra is the first Surface to ship this silicon. Partners launching RTX Spark laptops in fall 2026: ASUS, Dell, HP, Lenovo, MSI.

What this unlocks:

  • Local 120B inference with 1M-token context — no API calls, no per-token cost, no data egress
  • On-device fine-tuning for models that previously required dedicated cloud GPU instances
  • NVIDIA OpenShell — a runtime to securely host agents on the primary device
  • 2× faster Adobe Premiere & Photoshop — rearchitected from the ground up for the platform

For data-sovereign markets: For Saudi Arabia (PDPL, SDAIA), UAE, and Qatar specifically — where regulated data cannot leave-jurisdiction — a 128GB unified-memory laptop that runs frontier-class models on-device is a category shift. Plan procurement for late 2026 / early 2027.

06 · Vision + Agents: Qwen3.7-Plus — vision-grounded computer-use from Alibaba

JUNE 2 · 79.0 ScreenSpot Pro · 70.3 Terminal-Bench

Alibaba's Qwen team released Qwen3.7-Plus, the multimodal counterpart to the text-only Qwen3.7-Max. Built on the Max language backbone, Plus adds vision and GUI grounding — the ability to look at a screenshot and pinpoint exactly which pixels to click. Available now via Alibaba Cloud Model Studio (Bailian platform).

79.0 on ScreenSpot Pro places Qwen3.7-Plus at the front of the open-API GUI agent field. Plus is positioned as a hybrid GUI + CLI agent: the same model operates a browser through screenshots and runs shell commands in a terminal, with orchestration baked into the model rather than the framework.

Qwen3.7-Plus is explicitly designed to work across Claude Code, OpenClaw, Qwen Code, and other agent frameworks. For Arabic-language workloads specifically, the Qwen line continues to be the strongest non-Anthropic option.

07 · Hybrid Inference: Perplexity Computer — the data center moves to your machine

COMPUTEX · Hybrid local + cloud · Search as Code

Perplexity announced two major shifts in one week.

Hybrid Agentic Inference: Perplexity Computer can now split tasks between a local model on your device and frontier models in the cloud — automatically, task by task. The orchestrator reasons about which work needs the frontier and which can stay local, keeping sensitive data on-device.

Search as Code: The Perplexity Agent API now defaults to Search as Code — a new search architecture where the model writes Python that calls the search stack directly. Lower latency, fewer round trips, structured outputs by design.

The pattern: Between RTX Spark, Perplexity hybrid inference, and MiniMax M3's 1M context, the direction is unambiguous: the boundary between "local model" and "frontier model" is dissolving into a routing decision the orchestrator makes per-task.

08 · Coding: Copilot app, Codex Sites, Composer in Grok

GitHub Copilot App — expanded technical preview. An agent-native desktop experience built on GitHub. Go from issue to merged PR without leaving the app. This is the Copilot equivalent of Cursor: a dedicated environment for agentic coding, not a chat sidebar inside an IDE.

OpenAI Codex Sites — Codex can now turn work, ideas, and plans into an interactive website or app your team can explore and share with a URL. Rolling out first to Business and Enterprise plans. "Vibe coding" is no longer a meme — it's productized.

Composer 2.5 inside Grok Build — Cursor's Composer 2.5 is now available inside Grok Build. xAI's coding environment now offers the same Kimi K2.5-based agent at the same prices, expanding cost-efficient agentic coding beyond Cursor.

09 · Agents: Hermes Desktop, NVIDIA NemoClaw, and Higgsfield

Hermes Desktop — The Hermes Agent ecosystem now ships with a desktop application — taking the orchestration kanban out of n8n/web and onto the OS.

Hermes Agent + NVIDIA NemoClaw + OpenShell — The big partnership: NVIDIA × Hermes Skills Hub integration covering CUDA-X and Omniverse. Connect Hermes Agent with NVIDIA NemoClaw and OpenShell to Slack, Outlook, GitHub, and NVIDIA developer forums — then turn a chat correction into a reusable skill.

Claude + Higgsfield MCP — real estate marketing as a service — Higgsfield released a workflow that turns Claude into a real-estate marketing agency: analyze listings from Airbnb, Booking, Zillow, or Expedia → build the assets the owners never had (3D tours, clean site, photos that sell) → pitch the owner → scale to the next one. For GCC agencies, this maps directly to KSA short-term rental, Dubai property, and Doha hospitality.

Higgsfield plugins for Adobe Premiere & After Effects — The Higgsfield motion stack is now live as Premiere Pro and After Effects plugins. For agency post-production teams in Riyadh, Jeddah, and Dubai, this is the moment to bring AI motion design into the existing pipeline.

10 · Infra + OSS: Browser infrastructure, training on a laptop, and PewDiePie's open-source workspace

New Browser Infrastructure for agents — A new browser-infra layer dropped, purpose-built for agentic browsing: Chromium fork, Firecracker fork, custom Linux kernel. 3× cheaper than the previous best ($0.02/hour). Unlimited scaling. Subsecond cold starts. For any agent that needs to browse the web at scale — research agents, scraping, KYC verification, price monitoring — this is the new economic floor.

Training 120B+ models on a laptop — Unsloth + NVIDIA + Microsoft confirmed: you can now train 120B+ parameter models locally on the 128GB unified-memory RTX Spark laptop. Not inference — full training. For research teams in the GCC universities and Vision 2030-aligned AI labs, this is the largest single drop in research-iteration cost since LoRA.

PewDiePie launches Odysseus — a fully open-source, self-hosted AI workspace. Runs entirely on your own hardware. No cloud. No accounts. No telemetry. 110M+ subscribers being told "you can own this stack" is the largest privacy + sovereignty pitch of the year.

11 · Trending Repos: 10 GitHub repos that should not be free. But they are.

These ten projects replace what would otherwise be $1,000–$24,000/year in commercial tooling:

  1. The-Swarm-Corporation/AutoHedge (finance) — Four AI agents running an autonomous hedge fund on Solana. Director. Quant. Risk manager. Execution agent. pip install and start trading immediately.
  2. HKUDS/Vibe-Trading (finance) — 64 finance skills. 29 specialist agent swarms. Ichimoku, Elliott Wave, Black-Scholes, full Greeks. Watch agents debate strategies in real time.
  3. Fincept-Corporation/FinceptTerminal (finance) — Bloomberg Terminal replacement on your laptop. CFA L1–L3 analytics. 20+ investor AI agents including Buffett, Dalio, Soros personas. Bloomberg charges $24K/year. This is free.
  4. danny-avila/LibreChat (interface) — Every model ChatGPT runs, plus Claude, Gemini, DeepSeek, and 20 more. Self-hosted. Native MCP support.
  5. Anil-matcha/Open-Higgsfield-AI (creative) — 200+ AI models in one self-hosted interface. Flux, Midjourney, Sora, Kling, Veo, GPT-4o, SDXL. No subscription.
  6. Open-LLM-VTuber/Open-LLM-VTuber (personal) — Live2D AI companion that runs offline, sees your screen, hears your voice, and never forgets.
  7. AgriciDaniel/claude-ads (marketing) — Free Claude Code skill. 190 audit checks across Google, Meta, YouTube, LinkedIn, TikTok, Microsoft Ads. 6 parallel subagents. Agencies charge $4K/mo for this.
  8. cloudflare/agentic-inbox (email) — Cloudflare open-sourced an email client where an AI agent reads your inbox and drafts replies. Your email never leaves your account.
  9. jo-inc/camofox-browser (infra) — Open-source headless browser that makes AI agents invisible to bot detection. Accessibility-tree output drops token cost by 90%.
  10. heygen-com/hyperframes (video) — HeyGen open-sourced a video framework that renders MP4 from HTML. No React. No JSX. GSAP, Lottie, Three.js all work. Deterministic video generation.

The pattern: vertical agent skill packs are eating commercial SaaS. For agency procurement teams, the next budget review should re-evaluate every $300–$2,000/month tool against its OSS equivalent.

12 · By the Numbers: The state of the stack as of June 3, 2026

  • 120B — Local params on a laptop
  • 1M — Tokens · Open-weights context
  • 180K — OpenClaw GitHub stars
  • $0.02 — Per hour · Agent browser
  • 12× — MiniMax M3 vs Opus 4.8 cost
  • $24K — Bloomberg cost replaced · Free

13 · Bottom Line: What to do this week

  • For Claude teams: Upgrade to claude-opus-4-8. Pilot Dynamic Workflows on one large multi-step task and measure cost-per-task vs Opus 4.7.
  • For M365 enterprises: Start the Scout pilot conversation. Get the Frontier license, run a 30-day pilot, build the IT case before the year ends.
  • For infra teams: Plan RTX Spark procurement for Q4. 128GB unified memory + 120B local is the right baseline for AI dev teams in 2027.
  • For Conneqt Brain: When MiniMax M3 weights drop (~June 11), run your Arabic eval set. 1M context + open weights + frontier coding makes it a strong candidate for a self-hosted Brain backbone.
  • For agency ops: Install claude-ads + Adobe Higgsfield plugins. Two zero-cost upgrades this month.
  • For finance-curious: Stand up FinceptTerminal locally. CFA-grade analytics + 20+ investor agents + free.

The pattern to watch: Three rounds ago, the moat was the model. Two rounds ago, the moat was the post-training. This round, the moat is where the model runs. Microsoft made the enterprise agent OS layer. NVIDIA made the laptop big enough to be the cloud. MiniMax made the open weights large enough to be the frontier. Perplexity made the orchestrator smart enough to split between them. For teams building in the GCC, the strategic question for the next six months isn't "which model" — it's "where does each part of my workload run, and who governs it?"

— Conneqt