The AI industry shifted hard this week. Here's what matters for the GCC.
The first week of May 2026 was one of the busiest the industry has had all year. Anthropic struck a deal with SpaceX that almost no one saw coming, Code with Claude San Francisco shipped a wave of new agent capabilities, OpenAI replaced the default ChatGPT model, Google warmed up for I/O, and the open-source field continued to close the gap with the frontier labs.
This is the first of two rounds for May. Below: every story your team should be tracking, what was actually announced, and how it lands for businesses operating in Saudi Arabia and the wider GCC.
1. Anthropic Signs Compute Deal with SpaceX — Will Use All of Colossus 1
On May 6, 2026, Anthropic announced an agreement with SpaceX (now operating its AI division as SpaceXAI after its merger with xAI in February) to access the entirety of the Colossus 1 data center in Memphis, Tennessee.
The deal gives Anthropic more than 300 megawatts of new compute capacity — over 220,000 NVIDIA GPUs including H100, H200, and GB200 accelerators — coming online within the month. Anthropic also publicly stated interest in partnering with SpaceX on multiple gigawatts of orbital AI compute capacity, signalling that the future of frontier training may not be confined to Earth.
The immediate impact for users:
- Claude Code five-hour rate limits doubled across Pro, Max, Team, and seat-based Enterprise plans
- Peak-hours throttling on Pro and Max plans removed
- Claude Opus API rate limits raised "considerably"
The political subplot is hard to ignore. Elon Musk has spent the past year publicly criticising Anthropic, calling the company "misanthropic" and questioning its motives. On the day of the announcement Musk reversed course on X, saying he had spent time with senior Anthropic staff and was "impressed." He also added that SpaceX would "reserve the right to reclaim the compute if their AI engages in actions that harm humanity."
This deal joins Anthropic's other recent compute commitments: a 5GW agreement with Amazon, a 5GW partnership with Google and Broadcom landing in 2027, $30 billion of Azure capacity with Microsoft and NVIDIA, and a $50 billion U.S. infrastructure investment with Fluidstack.
Why it matters for the GCC: Capacity has been the single biggest bottleneck for businesses building serious workflows on Claude. With Claude Code limits doubled and Opus throughput up, the agencies and teams running Claude-based automations finally get headroom to ship.
2. Claude Launches Dreaming in Managed Agents — Plus Outcomes, Multi-Agent Orchestration, and Webhooks
The other big Anthropic announcement on May 6 came at Code with Claude San Francisco. Anthropic shipped four substantial upgrades to its Managed Agents platform:
Dreaming (research preview). A scheduled background process that reviews an agent's past sessions and memory stores, extracts patterns, and curates memory so agents improve over time. It surfaces recurring mistakes, workflows that agents converge on, and preferences shared across a team. Memory updates can happen automatically or be human-reviewed before they land. Dreaming is the missing half of the memory loop — memory captures what an agent learns mid-task, dreaming refines it between sessions.
Outcomes (public beta). You write a rubric describing what success looks like. The agent runs, and a separate grader — operating in its own context window so it isn't influenced by the agent's reasoning — evaluates the output against the rubric. If something isn't right, the grader pinpoints what needs to change and the agent takes another pass.
Multi-agent orchestration (public beta). A lead agent breaks complex jobs into pieces and delegates each one to a specialist with its own model, prompt, and tools. The example Anthropic gave: a lead agent runs an investigation while subagents fan out through deploy history, error logs, metrics, and support tickets, all working in parallel on a shared filesystem. Every step is traceable in the Claude Console.
Webhooks (public beta). Define an outcome, let the agent run, and get notified by webhook when it's done. This is what unlocks long-running, asynchronous agentic work in production.
Why it matters for the GCC: This is the most credible answer yet to "how do you actually run agents reliably for clients." For agencies productising agent systems for FMCG, legal, healthcare, and retail clients across KSA and the Gulf, multi-agent orchestration plus Outcomes-based grading is the architecture that finally takes agent work from demo to delivery.
3. Google Heads into I/O 2026 — Gemini 4.0 Expected May 19
Google I/O 2026 kicks off May 19 at the Shoreline Amphitheater in Mountain View, with the keynote at 10am PT (9pm KSA). The expectation across the industry: a major Gemini upgrade, widely rumoured to be Gemini 4.0, plus deeper integration across Workspace, Search, and Android.
In April, Google had already shipped:
- Workspace Intelligence — gives Gemini real-time context across Gmail, Chat, Calendar, and Drive without manual prompting
- Gemini Embedding 2 — a natively multimodal embedding model for text, image, video, and audio
- Gemini Enterprise Agent Platform — Google's unified platform to build, scale, govern, and optimize agents
- Gemma 4 — Google's new open-weight flagship in a 26B MoE configuration, running at roughly 85 tokens per second on consumer hardware
In the run-up to I/O, Google is pushing Gemini Canvas — a "vibe coding" environment where users describe an interactive project in natural language and the system generates it.
Why it matters for the GCC: Workspace Intelligence is the most underrated update of the quarter. Most businesses in the region run on Google Workspace; an AI layer that automatically pulls context from email, drive, and calendar without prompting changes how teams operate, with no workflow rebuild.
4. OpenAI Ships GPT-5.5 Instant — Now the Default Model
On May 5, OpenAI replaced GPT-5.3 Instant with GPT-5.5 Instant as the default model for every ChatGPT user, free and paid alike.
The headline improvements: more accurate answers in sensitive domains like medicine, law, and finance, fewer hallucinations, tighter responses with less filler, and improved STEM and image-understanding performance. The model also asks fewer unnecessary follow-up questions and generates fewer emojis and excessive headers.
For Plus and Pro users on web, GPT-5.5 Instant can now use context from past chats, files, and connected Gmail to make responses more personalized. Memory sources are now visible across all models so users can see where an answer came from, delete outdated sources, or correct mistakes.
GPT-5.5 Instant scores 81.2 on AIME 2025 (vs 65.4 for the older model) and 76 on MMMU-Pro (vs 69.2). Paid users keep access to GPT-5.3 Instant for three months before it retires.
OpenAI also confirmed that GPT-5.5, Codex, and Managed Agents are now available on Amazon Bedrock. Codex weekly active users have crossed 4 million.
Why it matters for the GCC: "Instant" defaults matter a lot. Most users — and most businesses — never change the model. A meaningful baseline upgrade flows directly into every tool and team using ChatGPT in the region.
5. Hermes Agent Updates — Now Self-Maintaining
Nous Research's Hermes Agent passed Claude Code in GitHub stars, crossing 100K in seven weeks — faster than LangChain or AutoGPT.
The April 30 release of v0.12.0 ("The Curator release") was the headline. Hermes now maintains itself: an autonomous background Curator grades, prunes, and consolidates the agent's skill library on its own schedule. Other shipped updates:
- Bundled Spotify and Google Meet integrations
- 18th and 19th messaging platforms (including Teams plugin)
- 4 new inference providers
- ComfyUI and TouchDesigner-MCP moved from optional to bundled by default
- Roughly 57% reduction in TUI cold start time
- A self-improvement loop that decides what's worth saving as a reusable skill
Hermes also introduced the Tool Gateway — paid Nous Portal subscribers now get web search, image generation, text-to-speech, and browser automation bundled with their subscription. No additional API keys.
The architecture is built around the agentskills.io open standard for skill portability. Hermes isn't asking the model to solve each task from scratch — it converts solved workflows into reusable skills, applies them to similar tasks next time, and refines them through use.
Why it matters for the GCC: For any team building productized agent services, Hermes' open architecture and self-maintaining skill library are an attractive alternative to closed platforms.
6. OpenClaw Ships Through a Heavy Security Cycle
The first week of May has been brutal for OpenClaw users on the security side. Multiple high and critical-severity CVEs were published between May 5 and May 6:
- CVE-2026-44109 (CVSS 9.8, critical) — authentication bypass via Feishu webhooks; affects versions before 2026.4.15
- CVE-2026-43534 (CVSS 9.1, critical) — input validation flaw escalating untrusted hook metadata to trusted system events; fixed in 2026.4.10
- CVE-2026-43530 (CVSS 8.8) — exec approval bypass via opaque multi-call binaries; fixed in 2026.4.12
- CVE-2026-44115 (CVSS 8.8) — shell expansion bypass via unquoted heredoc bodies; fixed in 2026.4.22
- CVE-2026-42434 (CVSS 8.8) — sandbox escape via host parameter override; fixed in 2026.4.10
- CVE-2026-44116 (CVSS 8.6) — SSRF in the Zalo plugin; fixed in 2026.4.22
- CVE-2026-42439 (CVSS 8.5) — SSRF policy bypass on browser tab actions; fixed in 2026.4.10
Why it matters for the GCC: Anyone running OpenClaw should be on 2026.4.22 or later. For internal multi-tenant deployments or anything touching client data, treat this as priority-one.
7. Notion Custom Agents Move to Paid Credits
Starting May 4, 2026, Notion's Custom Agents transitioned out of free trial into credit-based pricing at $10 per 1,000 Notion credits. Core AI features — Notion Agent, AI Meeting Notes, Enterprise Search — remain included on Business and Enterprise plans.
What's new in Notion 3.4 (April):
- AI Autofill with Custom Agent intelligence — continuously enriches, extracts, and categorises data across every database row
- Skills — turn "that thing you always ask the agent to do" into a one-tap workflow
- Calendar / Mail / Slack integrations for Notion Agent
- Salesforce and Box connectors
- AI Meeting Notes via API
- n8n MCP integration — connect Custom Agents to n8n automations
Why it matters for the GCC: For agencies and teams using Notion as their internal operating system, the move to credits means it's worth auditing which Custom Agents are actually saving time. The Slack, Salesforce, and Box connectors open the door to cross-tool agent work.
8. Obsidian 1.12.7 — CLI Becomes a First-Class Citizen
Obsidian's 1.12 series marks one of the biggest workflow shifts the app has ever made:
- Obsidian CLI — a real command line interface for scripting, automation, and integration with external tools
- Bases drag-and-drop import — dynamic database-style views over your notes
- Live Preview image resizing — corner-drag resize, double-click to reset
- iOS Share extension — save content from Safari and other apps directly into your vault
- iOS "Bookmark Link" Shortcut action
Why it matters for the GCC: For any team running an internal wiki or Second Brain on Obsidian, the CLI changes everything. You can now integrate Obsidian into automation pipelines — n8n, cron, custom Python — without hacks.
9. Open-Source Models — The Frontier Is Now Open
May 2026 begins with the most competitive open-weights field in AI history. Five frontier-class models have shipped in the last 30 days:
| Model | Params (Active) | License | Standout |
|---|---|---|---|
| DeepSeek V4-Pro | 1.6T / 49B | MIT | Strongest open coder, 1M context |
| DeepSeek V4-Flash | 284B / 13B | MIT | Cost-efficient, near V4-Pro perf |
| Qwen 3.5 | 397B / 17B | Apache 2.0 | 88.4% GPQA Diamond |
| Llama 4 Maverick/Scout | — | Meta custom | Scout: 10M-token context |
| Mistral Medium 3.5 | 128B dense | Apache 2.0 | 77.6% SWE-Bench |
| Gemma 4 | 26B / 4B | Apache 2.0 | 85 tok/s on consumer HW |
| Kimi K2.6 | 1T / 32B | Moonshot | Agent Swarm, 96.1% AIME |
| GLM-5.1 | 744B / 40B | Zhipu | Agentic deployment focus |
The pattern is clear: sparse Mixture-of-Experts is the default architecture, Apache 2.0 has won the license war, and context windows are no longer a meaningful gate.
Why it matters for the GCC: For regulated industries — healthcare, finance, government — open weights are the only path that satisfies data residency, PDPL compliance, and contractual data-handling requirements. Self-hosted Qwen 3.5 or DeepSeek V4 inside a VPC is now a credible alternative to API-based deployments.
10. Local Models — The Stack Just Got Serious
Running serious LLMs locally was a hobbyist activity in 2024. In May 2026 it's a deployment strategy.
Ollama remains the default CLI for developers. Single command install, OpenAI-compatible REST API at localhost:11434/v1, automatic GPU offloading. The OpenAI Python library works out of the box by overriding base_url.
LM Studio continues as the GUI-first option for non-technical users and model exploration.
Hardware tiers as of mid-2026:
- 8 GB VRAM (RTX 4060/4070): 7B models at Q4 quantization
- 24 GB VRAM (RTX 3090/4090): 4-bit models up to ~40B — DeepSeek-V2 21B, Qwen2.5 32B, Gemma 2 27B
- 48 GB+ (A100, dual 4090): 4-bit 70B-class models
- Apple Silicon: Mac Studio M-series 36GB+ is now a credible single-developer LLM workstation
Why it matters for the GCC: Local-first AI is no longer about saving money. It's about serving clients who can't send data to a US/EU API. The combination of Apple Silicon hardware, Ollama-style runtimes, and frontier-quality open-weight models means a small agency can now deliver enterprise-grade AI to a sovereignty-conscious client without a single byte leaving the country.
11. Manus — From Standalone Agent to Meta Property
Meta acquired Manus for roughly $2 billion in late 2025. On April 27, 2026, China's NDRC formally blocked the acquisition. Meta has said it will continue operating Manus technology in its products; Manus has said it will continue offering subscriptions from Singapore.
Recent Manus product updates:
- Web App Builder (March) — generates full websites and apps with built-in database, Stripe integration, and SEO
- Desktop app with local file access
- Manus integration in Meta Ads Manager (February 17) — free for Meta advertisers
- Wide Research, Slack integration, Mail Manus
Why it matters for the GCC: The Meta Ads Manager Manus integration is the most accessible entry point — free, infrastructure runs through Meta, and it covers high-volume use cases (report generation, audience analysis). Direct subscriptions to Manus.im should be evaluated carefully given the regulatory uncertainty.
12. Perplexity — Computer Goes Personal, Comet Goes Everywhere
Personal Computer — an always-on AI that runs on a dedicated Mac mini, merging local files, apps, and browsing sessions. Operates 24/7 as a digital proxy. Currently waitlisted.
Comet for iPad — proper native iPad support with multiple browser windows and Split View.
Comet for Android — launched with AI assistant, voice chat, and built-in ad blocking.
Comet Enterprise — AI-native browser for organizations. IT admins deploy via MDM, configure browser policies, and control which AI agent actions are permitted.
Numbers: 45M monthly active users, 170M global visitors monthly, $450M ARR as of March 2026, valued at approximately $20 billion.
Why it matters for the GCC: Comet Enterprise plus MDM deployment is the most underrated update of the spring. For regulated organizations wanting AI-powered browsing without personal-account-level employee access, this is the first viable path.
13. NVIDIA GTC 2026 — The Agentic Future Has Hardware
NVIDIA's GTC 2026 was less a chip launch and more a strategy reset. Jensen Huang's message: agentic AI requires a different stack.
Major announcements:
- Vera Rubin architecture — Combined GPU-HBM memory stacks memory directly on the chip, closing the compute/memory bottleneck
- Vera Rubin NVL72 — flagship rack-scale system, Microsoft Azure was first hyperscaler online
- Groq 3 LPX — 128 Groq LPUs per rack for inference; "35x higher throughput per megawatt"
- Vera CPU rack — 256 liquid-cooled Vera chips per rack; direct shot at Intel and AMD
- NemoClaw — enterprise OpenClaw with NVIDIA's stack, including OpenShell and a security sandbox
- DLSS 5 — neural shading with Bethesda, CAPCOM, and Ubisoft signed on
NVIDIA signaled $1 trillion in expected Blackwell and Vera Rubin orders by 2027. Data center revenue hit $193.5 billion in fiscal 2026.
Why it matters for the GCC: Inference cost is going to fall meaningfully as Vera Rubin capacity comes online. NemoClaw also creates a credible enterprise alternative for clients worried about the open-source security cycles in stock OpenClaw.
14. xAI — Now SpaceXAI
In May 2026, Elon Musk announced that xAI would cease to exist as a separate company. xAI and X are now consolidated under the SpaceXAI sub-brand of SpaceX, which acquired xAI in February at a combined valuation of $1.25 trillion.
Recent SpaceXAI / Grok announcements:
- Grok app integrations — May 6, deep integrations bringing apps directly into Grok
- Grok image generation upgrade — higher realism, stronger text rendering
- Grok Voice Cloning — April 30, clone a voice from a short recording
- Grok Voice Agent API — April 23, most capable voice agent now available to developers
- Grok Speech-to-Text and Text-to-Speech APIs — April 17, $0.10/hour batch, $0.20/hour streaming, 25+ languages
- Grok 4.20 Multi-agent Beta — available in the xAI Enterprise API
- Grok 5 roadmap — 7 models in training simultaneously, building toward 6T-10T parameters
Why it matters for the GCC: Grok is genuinely competitive on voice and image generation. For Arabic-language video, voice cloning, and creative pipelines, the new Grok APIs are worth real evaluation.
15. Railway — The Quiet Workhorse for Indie Deployments
Railway remains one of the most-used deployment options in the agentic AI ecosystem — especially for self-hosted OpenClaw, Hermes, and custom n8n deployments.
Key notes for May 2026: usage-based billing by the second, $5/month minimum on Hobby plan, volumes for persistent storage now functional, no built-in background workers. For OpenClaw specifically, Railway's one-click template is well-documented and container isolation reduces the risk profile.
Why it matters for the GCC: For agency-side deployments — internal automations, demo environments, client agent prototypes — Railway remains excellent. For client-production deployments handling real data, evaluate whether the reliability profile meets your SLA.
16. GitHub — Copilot Pricing Resets and Big Changes for Individual Plans
GitHub paused new sign-ups for Copilot Pro, Pro+, and Student plans, and is restructuring individual user billing.
Effective immediately:
- New sign-ups for Copilot Pro, Pro+, and Student are paused
- Tightened weekly and session usage limits
- Pro+ offers "more than 5x" the limits of Pro
- Opus models removed from Pro plans; Opus 4.7 remains on Pro+
- Refunds available through May 20 for users who cancel
Effective June 1, 2026:
- All Copilot usage moves to "GitHub AI Credits" — a token-based billing model
- Code completions and Next Edit Suggestions remain unmetered
- GPT-5.2 and GPT-5.2-Codex deprecated across Copilot
New product features: bring-your-own-model-key for Business/Enterprise (link your own keys for OpenRouter, Anthropic, OpenAI, or run local models via Ollama), inline diffs in chat, agents access any open terminal, integrated browser, and the /chronicle experimental command.
Why it matters for the GCC: For freelancers and small teams using Copilot Pro as a daily driver, the changes are disruptive. Bring-your-own-key on Business/Enterprise is the most interesting development — organizations can standardize on Copilot's UX while routing inference through their model of choice.
What Anthropic Calls Out for Builders
The week's compute and product news read together as a single message: the agentic era is past the prototype phase. Anthropic doubled limits because demand exceeded supply. Notion moved to credits because Custom Agents are doing real work. GitHub paused signups for the same reason. NVIDIA spent GTC pivoting from GPUs to a full agent stack.
For agencies and businesses in Saudi Arabia and the GCC, the practical takeaway is sharper. The infrastructure now exists to deliver enterprise agent systems. The models are good enough. The tools are mature. The bottleneck has moved from "can we build this?" to "can we operate this responsibly, in compliance, and at margins that work?"
That's the work for the next round.
— Conneqt