Vol.30 · May 11, 2026
dera news AI Weekly Vol.30 | 2026-05-11 - This Week's AI News
🤖 dera news AI Weekly Vol.30
Monday, May 11, 2026
This week's AI world in one sentence?
Anthropic secured large-scale compute from xAI/SpaceX, AISI confirmed Mythos can autonomously execute network compromises with the NSA already deploying it, and OpenAI shipped GPT-5.5 Instant to the free tier — AI's main battle moved on compute, capability, and reach all at once.
Anthropic locked in xAI/SpaceX's Colossus 1 data center (300MW, ~220,000 GPUs), then immediately doubled Claude Code's rate limits and dropped peak-hour throttling. The UK AI Security Institute (AISI) found Anthropic's Claude Mythos Preview can autonomously complete a 32-step corporate network compromise — and the NSA is already using it for vulnerability scanning, overriding Pentagon's "supply chain risk" warnings. OpenAI rolled out GPT-5.5 Instant to all ChatGPT users including the free tier, cutting hallucinations 52% on high-stakes prompts. China's DeepSeek is closing in on a $45B valuation led by a state-backed semiconductor fund. Open-source Hermes Agent topped OpenRouter's autonomous-agent benchmark, and Microsoft Agent Framework 1.0 went GA. Infrastructure, capability, reach, and capital all advanced in one week.
📊 What You Need to Know This Week
AI's Competitive Axis Splits Into Three: Compute, Capability, Reach
This week the AI competition broadened on every axis at once. Anthropic locked in 300MW of compute via xAI/SpaceX's Colossus 1 data center, then immediately doubled Claude Code's rate limits and dropped peak-hour throttling — securing compute now translates directly to product experience. At the same time, the UK AI Security Institute confirmed Anthropic's Mythos can autonomously execute a 32-step corporate network compromise, while the NSA is already running it for vulnerability scanning. And on the reach axis, OpenAI's GPT-5.5 Instant rollout to the ChatGPT free tier brought a 52% drop in hallucinated claims to every user, not just Pro subscribers.
What we're watching closely is competitors collaborating on compute when economics demand it, and frontier capability moving from "interesting" to "national-security inputs". Anthropic's 80× annual growth ran into a compute wall; xAI's Colossus 1 had spare GPU capacity to monetize before SpaceX's IPO — the economic logic overrode the public rivalry. Meanwhile Mythos's evaluation results signal that enterprise AI conversations have shifted from "should we use it" to "how deeply, and how do we defend against it."
💡 This Week's Actions
1. Actually try open-source Hermes Agent this week (1 hour) Hermes Agent just topped OpenRouter's autonomous-agent ranking. An open-source agent matching commercial performance is now installable and runnable on your own infra. The fastest way to know whether autonomous AI fits any of your workflows is to spend an hour: install it, run a minimal task, and your applicable scope becomes visible. → Hermes Agent details
2. Take advantage of doubled Claude Code rate limits (30 min) Anthropic doubled the 5-hour rate limits and removed peak-hour throttling. If you're already using Claude Code at work, extended uninterrupted coding sessions are now realistic. Restart the tasks you'd been throttling back, or take on the larger refactor or new feature you'd been deferring because of session caps. → Claude Code limits expansion
3. Re-evaluate ChatGPT for workloads you previously deferred (30 min) GPT-5.5 Instant cuts hallucinations 52% on high-stakes prompts (medicine, law, finance). Workflows you previously paused because "ChatGPT isn't reliable enough" may now cross your threshold. Pull up the list of internal "should we use it for X" questions you set aside, and re-evaluate this week. → GPT-5.5 Instant rollout details
📊 Monthly Deep Dive
Every month, we analyze the AI industry through 3 key shifts, action checklists, and editorial analysis.
👉 Read the latest monthly report
📰 This Week's AI Articles (All 14)
1️⃣ xAI and Anthropic Ink Major Data Center Deal
🏷️ Topic: Strategic Partnership
What Happened? Leading AI developer Anthropic has reportedly signed a deal to utilize xAI's "Colossus 1" data center. This partnership highlights the intense competition for high-performance computing resources essential for training and running advanced AI models. While specific terms were not disclosed, such agreements are becoming crucial for scaling AI capabilities amidst growing demand.
Our take What's notable here is the strategic alignment between two major AI players to secure essential infrastructure. This signals that access to massive compute power is a critical bottleneck, and companies are willing to partner, even with competitors, to overcome it. We read this as a clear indicator of the "compute wars" heating up.
2️⃣ Anthropic Doubles Claude Code Rate Limits, Drops Peak-Hour Throttling
🏷️ Topic: Product / Developer Tools
What Happened? On May 6, Anthropic doubled the 5-hour rate limits for Claude Code and removed peak-hour reduced limits entirely. The xAI/SpaceX compute deal mentioned in the lead — Colossus 1's 300MW and ~220,000 GPUs — is what makes this possible. Developers can now run extended coding sessions without throttling interruptions.
Our take This is a direct response to "we'd live in Claude Code if the rate limit weren't in the way." The broader takeaway: securing compute isn't just a business decision — it directly determines product experience. In a market where Copilot and Cursor compete on UX, sustained long-session availability becomes a meaningful differentiator.
3️⃣ AISI Evaluation: Mythos Demonstrates Autonomous Cyberattack Capability
🏷️ Topic: Ethics & Safety / Governance
What Happened? The UK AI Security Institute (AISI) found that Anthropic's "Claude Mythos Preview" can autonomously execute sophisticated cyberattacks previously thought beyond current AI. It completed a 32-step corporate network takeover scenario (TLO) and scored 73% on expert-level tasks. Government agencies, banks, and utility operators have since expressed opposition to a general public release of the model.
Our take The threshold shift here is what matters: AI can now autonomously complete network compromises end-to-end as the attacker. For SMEs that's a signal that the assumptions underlying current security posture are aging fast — both for adopting AI-augmented defense (e.g., Anthropic Claude Security) and for revisiting how internal LLM access is gated. Not a next-quarter project; a this-week project.
4️⃣ NSA Uses Anthropic's Mythos Despite Pentagon's "Supply Chain Risk" Warning
🏷️ Topic: Government / AI Adoption
What Happened? Axios reported on May 4 that the NSA is using Anthropic's latest model "Mythos Preview" for vulnerability scanning, even as the Department of Defense has flagged Anthropic as a "supply chain risk." The NSA made the call independently — the U.S. government does not have a unified position on which frontier AI labs are safe to use.
Our take The interesting shift is that frontier AI adoption in government has moved past "is this safe to use" to "which agency decides for itself, on what basis." For SMEs that means waiting for a government green-light on AI is increasingly disconnected from how actual procurement decisions get made. Each agency evaluates risk individually — your own organization probably will too.
5️⃣ OpenAI Ships GPT-5.5 Instant to the Free Tier — Hallucinations Down 52%
🏷️ Topic: LLM Development / Product
What Happened?
On May 5, OpenAI made GPT-5.5 Instant the default ChatGPT model for all users including the free tier. Compared to GPT-5.3 Instant, hallucinated claims dropped 52.5% on high-stakes prompts in medicine, law, and finance. Users can verify and manage the model's grounding, and the same model is available via API as chat-latest.
Our take The free-tier rollout is what changes the workplace calculus. Workloads previously held back because "ChatGPT is useful but not reliable" — verification-heavy tasks in regulated domains — flip when error rates halve. If your team has been deferring a "can we use ChatGPT for X" decision, this is the prompt to re-evaluate.
6️⃣ DeepSeek Closing in on $45B Valuation, Led by Chinese State Fund
🏷️ Topic: Investment / Chinese AI
What Happened? The Financial Times reported on May 6 that DeepSeek is in talks for a funding round valuing the company at roughly $45 billion, led by China's largest state-backed semiconductor investment vehicle. The valuation reflects DeepSeek's success in maintaining a ~6× cost gap below leading U.S. models while keeping near-frontier capability — and Beijing's evident interest in protecting that position at a national-strategy level.
Our take The scale of state backing makes this less of a startup story and more of a sovereign-AI story. DeepSeek's V4 Pro and Flash models remain genuinely useful for cost-sensitive workloads, but operating them now means folding supply-chain and geopolitical risk evaluation into model-selection criteria. U.S.–China hybrid deployments are looking less optional.
7️⃣ Hermes Agent Ascends to Top of AI Agent Rankings
🏷️ Topic: Open Source
What Happened? Hermes Agent, an open-source autonomous learning AI agent, has surpassed OpenClaw to claim the top spot in OpenRouter's global rankings. This achievement highlights the rapid progress in open-source AI agent development, demonstrating that powerful, self-learning AI capabilities are becoming increasingly accessible. Its strong performance suggests significant potential for adoption by smaller businesses and individual developers.
Our take Our take is that Hermes Agent's rise to the top is a testament to the power of the open-source community in pushing AI boundaries. It signals that high-performing, autonomous AI is not exclusive to large corporations, making advanced capabilities more widely available. This is exciting for fostering innovation and diverse applications.
8️⃣ Anthropic Partners with Financial Giants for Enterprise AI Venture
🏷️ Topic: Strategic Partnership
What Happened? Anthropic has launched a new enterprise AI business initiative in partnership with major financial firms like Blackstone and Goldman Sachs. This collaboration aims to accelerate the adoption of Anthropic's Claude AI model within the financial sector, leveraging the expertise and client networks of these institutions to drive broader market penetration and intensify competition with rivals like OpenAI.
Our take What's notable here is Anthropic's focused strategy on specific high-value enterprise verticals. This signals a mature approach to market entry, recognizing that deep industry partnerships can unlock significant adoption faster than a purely generalist approach. We're watching how this specialization impacts their competitive standing against more generalized LLM offerings.
9️⃣ SpaceX Reportedly Plans Semiconductor Plant in Texas
🏷️ Topic: Hardware
What Happened? Building on the trend above, Elon Musk's aerospace company, SpaceX, is reportedly moving forward with plans to construct a semiconductor manufacturing plant, potentially named "Terafab," in Texas. This ambitious project could involve an initial investment of up to $55 billion, aiming to produce chips crucial for both SpaceX's space ventures and xAI's artificial intelligence initiatives, securing a vital component of their tech ecosystem.
Our take This move by SpaceX is a powerful signal of vertical integration, extending beyond software to the very silicon layer. Our take is that this isn't just about securing supply; it's about gaining a competitive edge by controlling the entire hardware stack, reducing reliance on external suppliers, and potentially innovating chip design specifically for AI and space applications.
🔟 Microsoft Agent Framework 1.0 Arrives, Empowering Multi-Agent AI
🏷️ Topic: LLM Dev
What Happened? Microsoft's Agent Framework has reached version 1.0, marking its transition to a stable release. This framework provides a robust foundation for developing complex AI agents, including multi-agent systems that can collaborate to achieve sophisticated goals. Its stability now opens the door for small and medium-sized businesses to more readily implement advanced AI automation solutions.
Our take What's notable about this release is the emphasis on stability and accessibility. By reaching 1.0, Microsoft is signaling that multi-agent systems are ready for broader enterprise adoption, not just research labs. We read this as a significant step in democratizing the ability to build sophisticated, task-oriented AI workflows.
11. GitHub Unveils Spec-Kit to Enhance AI Code Generation
🏷️ Topic: LLM Dev
What Happened? GitHub has released Spec-Kit, an open-source tool designed to address the challenge of ambiguous instructions in AI-assisted code generation. Spec-Kit aims to prevent misinterpretations by providing clearer specifications, enabling AI models to generate higher-quality code directly from detailed requirements, thereby streamlining the development workflow and reducing rework.
Our take Our take is that Spec-Kit is a practical, much-needed step towards making AI code generation truly reliable. It acknowledges the "garbage in, garbage out" problem even with powerful AI, and by focusing on better inputs, it helps developers leverage AI more effectively. This signals a maturing understanding of how humans and AI can best collaborate in coding.
12. Meta AI Unveils NeuralBench for Brain-Inspired AI Benchmarking
🏷️ Topic: LLM Dev
What Happened? The Meta AI team has introduced NeuralBench, an open-source framework designed to standardize the evaluation of neuro-inspired AI models. This platform aims to provide a consistent and comprehensive way to benchmark models that draw inspiration from brain structures and functions, which is expected to accelerate research and development in this specialized field of AI.
Our take What's notable here is Meta's commitment to standardizing evaluation in a niche but potentially transformative area of AI. By open-sourcing NeuralBench, they are fostering collaboration and transparent progress in brain-inspired AI. We read this as an investment in fundamental research that could yield significant long-term breakthroughs.
13. ARIS: New Autonomous AI Research Tool Emerges
🏷️ Topic: LLM Dev
What Happened? Shanghai Jiao Tong University has released ARIS, an open-source autonomous AI research tool. ARIS is designed to facilitate collaborative research by enabling multiple AI agents to work together, critically evaluating each other's findings to achieve more reliable and long-term research outcomes. This approach aims to enhance the trustworthiness and consistency of AI-driven scientific discovery.
Our take Our take is that ARIS represents an intriguing approach to AI-assisted research, emphasizing collaborative scrutiny among AI agents. This signals a move beyond single-agent problem-solving towards more robust, self-correcting research paradigms. We're watching how this "AI peer review" concept influences the quality and pace of scientific discovery.
14. StateSMix: A Novel Mamba-Powered Compression Technique
🏷️ Topic: LLM Dev
What Happened? A new data compression technology called StateSMix has been unveiled, leveraging the efficient Mamba model. This innovative technique operates without requiring GPUs or extensive pre-training, demonstrating superior efficiency compared to existing compression methods. Its ability to achieve high performance with minimal hardware demands makes it a promising solution for various data-intensive applications.
Our take What's notable about StateSMix is its practical application of the Mamba model to a fundamental problem like data compression, without the typical heavy compute requirements. This signals a broader trend of optimizing AI models for efficiency and accessibility, making advanced techniques viable even in resource-constrained environments. We're watching its potential impact on edge computing and data transfer.
📚 Editor's Note
What stood out this week for the editorial team was the feeling that AI has fully left the "software technology" phase behind. Anthropic securing 300MW from xAI/SpaceX and immediately opening up Claude Code; Mythos autonomously completing network compromises with the NSA already deploying it; GPT-5.5 Instant rolling out to the free tier with hallucinations cut in half — none of this fits "AI is a useful tool" anymore. Compute, capability, and influence are all rewriting societal defaults at once.
For SME decision-makers, the conversation has moved past "should we adopt AI" into "where do we use it, how do we defend, which supply chain do we depend on." GPT-5.5 Instant lowered the threshold for business adoption. Mythos changed defensive assumptions. DeepSeek's state-backed round made supply chain a procurement criterion. The number of axes you have to reason about just went up.
We'll be back next week with useful information and food for thought.
dera news Editorial Team
🤝 Ready for Team-Wide AI Adoption?
Once you've tested ChatGPT individually and felt its potential, the next step is team-wide implementation.
But where do you start? Which tools to choose? How to roll out internally? How to measure ROI?
Let's talk about your specific challenges and goals.
We'll discuss your business needs and propose the optimal implementation roadmap for your organization.
📬 About this newsletter