The AI news cycle is dominated today by the widening consequences of the U.S. export ban on Anthropic's frontier models, which is spurring a wave of homegrown rivals across Asia. Elsewhere, the money keeps flowing into infrastructure as Big Tech's data center commitments push past $850 billion, enterprises keep raising AI budgets, and identity vendors race to bring governance to autonomous agents. There is also notable progress at the edges of the field: Meta demonstrated non-invasive brain-to-text decoding, Google added scientific models for drug discovery, and California struck a discount deal to put Claude in the hands of government workers. Below is a synthesis of the day's most significant developments.
Model and Product News
The export ban on Anthropic's Mythos line continues to reshape the global model market. Cognition released Devin Fusion, a multi-model harness that routes work between frontier and cheaper models using a dual-agent design — a main agent paired with a "sidekick" for dynamic routing — cutting costs by roughly 35% on the FrontierCode benchmark while holding top-tier performance and avoiding expensive cache misses. In Asia, the restrictions are accelerating domestic alternatives. Tokyo-based Sakana AI launched Fugu, an agent-capable frontier model; its Fugu Ultra tier reportedly scored 93.2 on LiveCodeBench, edging out comparable models, and starts at $5 per million input tokens. Separately, Chinese cybersecurity firm 360 unveiled Tulongfeng, alongside a tool called Yitianzhen, positioned for cyber defense and incident response as a rival to Anthropic's offering.
Efficiency was a recurring theme. DeepSeek introduced DSpark, a system that speeds up model responses without altering their meaning by running a lightweight "scout" a few steps ahead to guess the likely output path, letting the larger model quickly verify safe steps. Google expanded its consumer and enterprise reach on two fronts: it made its personalized image-generation feature free for eligible U.S. users, generating images tuned to a user's inferred preferences through an opt-in personalization layer, and it began offering SandboxAQ's large quantitative models through Google Cloud for drug discovery, materials science, and semiconductor work. Mistral, meanwhile, launched Workflows, a durable, fault-tolerant orchestration platform for running and monitoring multi-agent pipelines.
Enterprise AI and Infrastructure
The capital intensity of the AI buildout reached a new marker, with major cloud and tech firms now committing more than $850 billion in future data center lease obligations. Meta and Microsoft posted especially large increases, underscoring how aggressively hyperscalers are locking up compute. Demand on the buyer side looks equally durable: a new survey of CIOs and technology leaders found enterprise AI spending still accelerating, with most organizations treating AI as net-new budget rather than a replacement for existing software, and OpenAI remaining the most-used provider. That growth is straining governance, prompting calls for cost controls that tie AI usage to business outcomes rather than raw token spend.
Governance of autonomous agents is quickly becoming its own product category. Okta made "Okta for AI Agents – Core" generally available for FedRAMP- and HIPAA-regulated environments, registering agents as first-class identities with least-privilege, short-lived scoped tokens, continuous audit logging and SIEM streaming, and a kill switch to contain agents that stray off-mission. AWS added watermarks to EC2 machine images so organizations can trace approved images across copies, regions, and shared accounts. On the funding side, a startup called Engram raised $98 million — with backing from General Catalyst, Kleiner Perkins, Sequoia, and Andrej Karpathy — to reduce the token usage and cost of running AI at scale.
AI Coding and Agents
Agentic development tools kept maturing. Cursor launched a public iOS beta that lets developers launch and control cloud or local coding agents from a phone, receive updates through live activities, and merge pull requests on the go — following similar mobile moves by Anthropic and OpenAI. OpenAI teased a hardware device for Codex users, a compact shortcut keyboard built with keyboard maker Work Louder, with a fuller reveal slated for mid-July. The analytical commentary was more sober: several writers argued that AI has shifted the software bottleneck from writing code to deciding what to build, meaning teams need more product thinking and stronger fundamentals rather than blind trust in generated output. Related research suggested that while assistants speed up code creation, review, testing, and delivery are becoming the new constraints, and a new benchmark, RoadmapBench, was introduced to test agents on long-horizon, multi-file coding tasks grounded in real repository upgrades.
Healthcare, Science, and Government
Meta showcased Brain2Qwerty v2, a system that converts brain activity into written text without an implant. It uses magnetoencephalography to read signals from outside the head, was trained on recordings from nine volunteers who each spent about ten hours typing inside an MEG scanner, and achieved 61% average word accuracy — 78% for the best participant — a meaningful gain for non-invasive brain-to-text. Meta framed the work as aimed at helping people who have lost the ability to communicate. In the public sector, California Governor Gavin Newsom announced a deal with Anthropic giving state agencies and local governments access to Claude at a 50% discount, bundled with training and support, to help public employees draft documents and analyze information.
Quick Takes
A new RBC-style enterprise survey reaffirmed that AI budgets are being treated as incremental spend, not substitution, with OpenAI leading provider usage.
"The Economy of Tokens" analysis argued AI is shifting from closed, vertically integrated systems toward more open, modular markets.
A widely shared piece on reinforcement learning "beyond the verifiable" examined how to extend RL's gains from checkable domains to fuzzier tasks, and which companies are attacking the problem.
DiScoFormer, a single transformer handling both density and score estimation across distributions, drew attention in the research community.
Salesforce's promotion of "Claude Tag" inside Slack caused internal confusion, given Slack's own Slackbot and Agentforce; Salesforce expects to spend roughly $300 million on Anthropic tokens this year and holds about a 1% stake in the company.
Google Cloud proposed an Open Knowledge Format, an open specification for packaging metadata and curated context for both humans and AI systems.
Analysts continued debating whether AI creates a single shifting bottleneck in product development or many, cautioning against overconfident predictions about how work will change.
A consumer-oriented item circulated a set of ChatGPT prompts for evaluating expensive purchases — comparing alternatives and surfacing hidden downsides before buying.
What This Means for Your Business
The clearest signal for buyers is that model choice is becoming a portfolio decision rather than a single-vendor bet. Tools like Devin Fusion and cost-reduction startups such as Engram exist because routing work across a mix of frontier and cheaper models is now the practical path to controlling spend. Small and mid-sized businesses can apply the same logic without building their own harness: reserve premium models for genuinely hard tasks and route routine work to cheaper ones. As Asian labs release capable, lower-priced alternatives, the competitive pressure should continue to push per-token prices down, which favors patient buyers who avoid locking into long, inflexible commitments.
The governance story deserves attention even at companies that do not think of themselves as heavy AI users. The launch of agent identity platforms with least-privilege tokens, audit trails, and kill switches reflects a maturing recognition that autonomous agents need the same access controls as employees — and often stricter ones. Before deploying agents that can touch customer data, financial systems, or production infrastructure, businesses should decide how those agents authenticate, what they are permitted to do, and how their actions are logged. That discipline is far cheaper to establish now than to retrofit after an incident.
The surveys showing AI budgets treated as net-new spend are a double-edged data point. They confirm that adoption is real and durable, but they also warn that costs can sprawl across departments faster than finance can track them. The recurring advice — measure AI against business outcomes rather than usage volume — is worth adopting early. For a small business, that might mean tying an AI investment to a concrete metric like faster quote turnaround or fewer support escalations, rather than counting seats or tokens, so the spend can be justified or cut on evidence.
For teams building software, the shift in where value is created is the practical takeaway. If assistants are compressing the cost of writing code, the differentiators become deciding what to build and verifying that it works. Businesses hiring or reskilling should weight product judgment, domain knowledge, and review capacity more heavily, and invest in testing and delivery pipelines rather than assuming faster code equals faster shipping.
Finally, developments like Meta's non-invasive brain-to-text work, Google's scientific models, and California's discounted access to Claude point to AI moving into higher-stakes, more regulated domains — healthcare, scientific research, and public services. For businesses in or adjacent to those sectors, the opportunity is real but comes with heightened expectations around accuracy, privacy, and accountability. Following the emerging governance playbook — auditable outputs, clear human oversight, and provenance for AI-assisted decisions — will increasingly be a precondition for operating in these markets rather than a nice-to-have.