This Week in AI is an AI-generated weekly roundup, curated and reviewed by the Kursol team. We use AI tools to gather, summarise, and analyse the week's most important developments — then add our perspective on what it means for your business.
Anthropic cut the price of Claude Sonnet 5 by 50% this week, while separately disclosing that Alibaba ran 29 million conversations through the model to steal it. Meanwhile, the infrastructure arms race hit a new milestone: 175 MW of operational AI compute came online. This mix of capability gains, competitive threats, and capacity milestones tells you everything you need to know about where enterprise AI investment is heading right now.
Anthropic Launches Claude Sonnet 5 at Half the Expected Price
Anthropic released Claude Sonnet 5 on June 30, positioning it as the "most agentic Sonnet yet"—meaning it can run autonomous tasks, use tools (like browsers and terminals), and reason through multi-step problems that previously required the more expensive Opus 4.8.
The key business angle: pricing. Sonnet 5 launches at $2 per million input tokens and $10 per million output tokens through August 31, undercutting what observers expected. After August, it moves to $3 and $15—still well below Opus 4.8's $15 and $75 pricing. The model became the default for all Free and Pro users starting July 1, meaning any team already on Anthropic's platform gets the upgrade automatically.
Performance is close to Opus 4.8 on reasoning, coding, and knowledge work—the exact tasks mid-market and enterprise teams rely on most. The agentic capabilities matter because they mean less hand-holding from your team. Sonnet 5 can make plans, run loops, and fix its own mistakes rather than asking for human input between steps.
Anthropic also restored access to its Fable and Mythos frontier models after an 18-day pause caused by a US government export control review. Fable 5 is now back in production.
Why it matters for your business: If your team is evaluating Claude for new projects, the pricing shift changes the maths. What was a $0.75 per 1,000 input tokens (Opus 4.8) is now $0.02 per 1,000 input tokens (Sonnet 5 intro pricing)—roughly 37x cheaper for the same reasoning capability on most real-world tasks. That maths inverts the ROI calculation on agentic workflows. You can now experiment with autonomous agents at a cost that was previously untenable for mid-market teams. The automated reasoning, tool use, and task planning means less overhead managing handoffs between your team and the AI.
This is the kind of vendor move that makes teams like yours suddenly able to afford AI automation use cases that didn't make sense six months ago. Teams that were waiting on pricing are now running pilots.
Anthropic Discloses Alibaba Stole 28.8 Million Claude Conversations
On June 24, Anthropic disclosed that operators tied to Alibaba's Qwen lab ran 28.8 million conversations with Claude between April 22 and June 5 using roughly 25,000 fraudulent accounts. The goal: extract Claude's reasoning and coding capabilities to train a competing model—a technique called model distillation.
This is not a traditional security breach. Alibaba never accessed Anthropic's source code, weights, or training data. Instead, they asked Claude millions of carefully chosen questions, recorded the answers, and fed those conversations into their own model to teach it to think like Claude. The focus was on areas where Claude excels: agentic reasoning, software engineering, and long-horizon task planning.
Anthropic sent a dossier to the US Senate Banking Committee on June 10. The scale is historic for Anthropic: it's nearly double the distillation campaign the company attributed to DeepSeek, Moonshot AI, and MiniMax combined in February (when those three labs were caught performing 150,000 to 13 million exchanges each).
Why it matters for your business: Model distillation is now a documented, industrial-scale attack vector. If you're choosing between Claude and a competing model built by a Chinese AI lab, you need to ask whether that model was trained partly on extracted Claude conversations. The disclosure doesn't change what Anthropic's API can do—Claude still performs the same tasks—but it should force enterprises to think about model provenance.
More broadly, this is a cautionary tale about cost arbitrage in AI. Alibaba's attack worked because Claude's API is accessible and relatively cheap. They could run millions of conversations at a cost that made sense for them. If your team relies on any third-party AI model, assume competitors are running distillation campaigns against it. That's a reason to diversify your models, version-lock your critical workflows, and monitor API usage patterns for suspicious activity.
The incident also raises a question many CFOs are asking: if frontier models can be distilled, what's the long-term competitive advantage? The answer is speed to capability, not permanence. That's another reason to focus on evaluating AI on the basis of workflow fit, not vendor lock-in.
Applied Digital Brings 175 MW Online at Polaris Forge 1
Applied Digital announced on July 1 that it achieved Ready for Service for Phase 1 of Building 2 at Polaris Forge 1, delivering 75 MW of operational AI compute capacity and bringing the total live capacity at the campus to 175 MW. The facility is contracted to deliver 400 MW at full buildout.
Context: Polaris Forge 1 is a hyperscaler AI data centre campus in North Dakota designed specifically for training and inference workloads. Applied Digital built it to run the kind of compute-intensive work that the major AI labs (OpenAI, Anthropic, Google, Meta) and cloud vendors (AWS, Azure, GCP) need. The 175 MW milestone means that capacity is now live and customer-facing.
The broader infrastructure race is accelerating. NVIDIA, chip startups, cloud vendors, and colocation operators are all racing to add power, cooling, and networking to feed the AI model training pipeline. Applied Digital's buildout is one of the largest capacity additions of the year.
Why it matters for your business: If your organisation trains custom models or runs inference at scale, Applied Digital's capacity becomes available to you via their customer (likely a major cloud vendor or AI lab). But more importantly, this signals that frontier model training capacity is becoming less of a bottleneck. When 175 MW of dedicated AI compute goes live in one quarter, it means the price of model training is heading down, and the pace of new model releases will accelerate. That's both an opportunity and a threat: more capable models arrive faster, but your competitive window to adopt them shrinks.
For mid-market teams, the second-order effect is cheaper fine-tuning and custom model work. As infrastructure costs drop, the economics of building a proof of concept on your own data (rather than using an off-the-shelf API) shift in your favour.
CREATE AI Act Advances Towards Permanent Status
A bipartisan group of senators is pushing the CREATE AI Act to put the National Artificial Intelligence Research Resource (NAIRR) on permanent federal footing. The bill, led by Sens. Martin Heinrich (D-NM), Todd Young (R-IN), Mike Rounds (R-SD), and Cory Booker (D-NJ), would establish NAIRR as a congressionally authorised, permanently funded infrastructure within the National Science Foundation.
A pilot version of NAIRR has operated since early 2024, supporting more than 600 cutting-edge AI research projects across all 50 states. The CREATE AI Act would codify it, ensuring that researchers and companies have sustained access to computing infrastructure, datasets, and tools for AI experimentation regardless of political winds.
Why it matters for your business: NAIRR is designed to democratise access to the compute and data infrastructure that historically only well-funded labs could afford. If your team wants to collaborate with academic researchers, build proprietary models on public datasets, or tap into a shared research infrastructure, NAIRR is the mechanism. Permanent funding means that access becomes reliable, not dependent on annual appropriations or executive orders.
For scaling companies in research-intensive domains (biotech, materials science, energy), NAIRR access is a competitive edge. If the bill passes, it's another tool in the kit for teams that want to experiment with AI without building entire infrastructure stacks in-house.
Quick Hits: More AI News This Week
Five Eyes intelligence alliance warns frontier AI timelines are collapsing: The UK, US, Canada, Australia, and New Zealand jointly stated that frontier AI models will exceed current expectations—and the timeline is "months, not years." This isn't hype; it's a direct signal that capabilities are accelerating faster than government can regulate or plan around.
Meta expands cloud partnerships for AI workload scaling: Meta announced expanded partnerships with cloud providers to help enterprises scale AI workloads. The move signals Meta's recognition that AI model deployment requires ecosystem partnerships, not just API access.
California's Poppy AI tool rolling out statewide: California's AI-powered tool for government operations, piloted with 2,800 employees across 67 state departments, is moving towards full statewide rollout. Early signal that government AI adoption is real and scaling.
What This Means for Your Business
Three distinct dynamics emerged this week, and they're worth sitting with for a moment.
First: Pricing pressure is real, and it favours builders over incumbents. Anthropic's Sonnet 5 pricing cuts and the infrastructure capacity growth both point to a market where the cost of AI is dropping faster than most enterprises expected. This doesn't mean AI gets cheaper indefinitely—frontier models will always cost more—but it means that the middle tier, where most business work happens, is seeing dramatic cost deflation. Teams that were shelving AI pilots because of pricing can now resurrect them. That's a competitive advantage for first movers.
Second: Model security is a new category of operational risk. The Alibaba disclosure isn't just about one company extracting Claude conversations. It's a signal that model distillation is an industrialised attack vector. If you're buying AI capabilities from any vendor—especially a closed vendor like Anthropic or OpenAI—you need to ask whether that model could be or has been distilled. Open-source models inherit different risks. The takeaway: diversify your model portfolio, version-lock critical workflows, and don't assume vendor durability.
Third: Infrastructure capacity is becoming a commodity, not a constraint. When 175 MW goes live in a quarter, the bottleneck for AI capability shifts from power and cooling to talent and data. Your constraint now is whether your team can find, clean, and structure the data to fine-tune models and run inference at scale. That's a recruiting and operations problem, not a hardware problem. This is exactly where external partners like Kursol help teams that don't have dedicated AI engineering capacity.
The gap between AI-ready and AI-late is widening every week. If you're unsure where your organisation stands, take our free AI readiness assessment to find out.
This Week in AI is Kursol's weekly analysis of the most important artificial intelligence developments — focused on what actually matters for your business. Subscribe to our RSS feed to never miss an edition.
FAQ
Sonnet 5 costs roughly 37 times less per input token than Claude Opus 4.8, with performance close enough for most reasoning, coding, and knowledge work. If your team was evaluating Claude but rejected it on cost grounds, the maths have shifted. You can now run agentic workflows—tasks that previously required expensive Opus instances—at Sonnet pricing. Run a new cost-benefit analysis on your use cases.
No, but it's a reason to diversify. Model distillation doesn't mean Claude is compromised or less capable. It means that competitive AI labs are reverse-engineering frontier models at scale. If you rely on a single vendor's model for critical workflows, you should add a second model (open-source or from a different vendor) to your toolkit. Version-lock your model deployments so you can switch quickly if a model becomes deprecated or compromised.
In the short term, your team doesn't feel it directly. But it means frontier model training capacity is expanding rapidly, which drives down model development costs. That translates to faster model release cycles and cheaper fine-tuning services. For your team, watch for new model announcements and declining prices on custom model training in 6-9 months.
Yes, if your team wants to experiment with AI without building your own infrastructure. NAIRR gives non-hyperscaler teams access to compute, datasets, and tools. If you're in biotech, energy, materials science, or any field where proprietary data + cutting-edge models = competitive advantage, NAIRR is worth exploring. It's especially valuable for mid-market companies that want to build custom models but can't justify the capital for a training infrastructure.
Very. Intelligence agencies don't issue joint statements lightly. This is a signal that AI capabilities are advancing on a faster timeline than government expected. For your team, it means staying current on new models becomes a monthly task, not a quarterly one. Your AI governance and vendor evaluation processes need to assume that frontier capabilities refresh faster than your procurement cycle.
Ready to get your time back?
No pitch, just a conversation about what Autopilot looks like for your business.