The AI Brief

Commerce Secretary Howard Lutnick wrote to Anthropic chief compute officer Tom Brown on June 26, stating that "appropriate safeguards are in place to permit certain trusted partners to access the Claude Mythos 5 Model." The letter, seen by Bloomberg and Axios, cited "significant progress" in daily talks between Anthropic and the government since the June 12 export control directive. More than 100 companies and federal agencies, described by a source cited by Reuters as including many Fortune 500 firms, received access. Many are part of Project Glasswing, Anthropic's invite-only cybersecurity program. The letter is silent on Fable 5, the public-facing version of the same model family, and Lutnick reserved the right to revoke access "should circumstances change."

On the same day, OpenAI previewed GPT-5.6 in three tiers: Sol (flagship), Terra (GPT-5.5-class performance at half the price), and Luna (high-volume, lowest cost). The launch reached approximately 20 government-approved organizations via the API and Codex. OpenAI stated it previewed the models with the administration ahead of release and launched with a restricted cohort "at their request," while publicly noting that customer-by-customer government approval "should not become the long-term default." OpenAI's system card classifies all three GPT-5.6 tiers at a "High" cyber-risk rating, while placing them below the "Cyber Critical" threshold in its Preparedness Framework. An August deadline under the June 2 executive order requires agencies to formalize a classified model assessment process for "covered frontier models."

The two moves together establish a structural precedent. The Trump administration has now demonstrated it can remove frontier models from the market on national security grounds, negotiate a resolution on its own timeline, and dictate which organizations receive restored access, all without invoking a formal statutory mechanism. Anthropic confirmed it is "working with the government to expand access to Mythos 5 and make Fable 5 available for general use again." The first covered version of this story ran in No. 14 (June 14, 2026).

Disclosure: Anthropic, mentioned throughout this item, is the company that develops Claude, which generates this brief.

Gartner published the $206.5B AI agent software figure in a May 5, 2026 press release covering autonomous business trends. The figure covers purpose-built agentic AI software, a narrower category than broader "AI software," and is projected to rise further to $376.3B in 2027. The baseline was $86.4B in 2025. Gartner separately found, in the same release, that approximately 80% of organizations surveyed report workforce reductions tied to AI, though those reductions are not yet translating reliably into measurable ROI (see Sleeper). The survey covered 350 global business executives at organizations with at least $1B in annual revenue that had deployed or piloted AI agents, intelligent automation, or related technologies. Gartner projects that more than 40% of agentic AI projects will be cancelled by end of 2027, which the firm attributes to escalating costs, unclear business value, and inadequate risk controls. Only 17% of organizations have deployed AI agents to date despite the spending surge.

Google published the ARD specification on June 17, 2026 under Apache 2.0, developed with launch partners including Microsoft, Hugging Face, Cisco, Databricks, GitHub, GoDaddy, Nvidia, Salesforce, ServiceNow, and Snowflake. The spec defines how organizations publish a machine-readable catalog (an ai-catalog.json file) on their own domain, which federated discovery services then index. An agent querying that registry gets back ranked matches with endpoint definitions, security scopes, and revocation URLs. GitHub simultaneously launched Agent Finder, a capability letting GitHub Copilot discover and call MCP servers and agents dynamically at runtime using ARD.

The spec is version 0.9 and in active draft; security researchers have flagged that discovery acting as the initial trust boundary introduces poisoned-registry and forged-catalog attack vectors that the current metadata-verification approach only partially mitigates. One analysis notes ARD standardizes one of five discovery layers; the remaining four, where most tool selection currently happens via model priors and system prompt injection, have no equivalent standard yet. An agent operating on a training cutoff from months prior may not fire an ARD registry query at all if it already "knows" a tool from pretraining data.

CNBC reported June 26 that companies that once prioritized frontier model access without cost controls are now demanding ROI justification and tighter budgets. Lindy CEO Flo Crivello told CNBC his startup switched 100% of Claude traffic to DeepSeek after AI costs became unsustainable, estimating savings of millions within months while still spending more on AI than on payroll. Uber burned through its entire 2026 AI budget in four months after Claude Code adoption in its engineering org rose from 32% to 84% between December 2025 and March 2026; monthly API costs per engineer ran between $500 and $2,000. Walmart introduced per-employee token caps after its global CTO cited the cost of running identical requests through AI agents versus traditional search. Microsoft has asked thousands of engineers to transition from Claude Code to an internally developed alternative by end of June 2026, widely attributed to cost management.

D.A. Davidson analyst Gil Luria told CNBC that current growth rates for both Anthropic and OpenAI are "the fastest they will ever be, which is mostly a matter of basic math," and named concern over enterprise token-spend limits as a reason both labs have incentive to go public now. Anthropic's annualized revenue run rate was $47 billion in May 2026; OpenAI's was tracking closer to $25 billion earlier this year. OpenAI CEO Sam Altman acknowledged to CNBC that ROI questions are "the most fair criticism right now of AI" and that cost concerns had gone from never coming up to the second-most-common issue he hears from customers in a matter of months. Gartner analysis finds agentic workflows require 5 to 30 times more tokens per task than a standard chatbot query, meaning that the shift from chatbot to agent deployment is the multiplier driving invoice shock, not any change in unit pricing.

Gartner Distinguished VP Analyst Helen Poitevin presented the finding at Gartner IT Symposium in May 2026: approximately 80% of organizations report workforce reductions tied to AI deployment, but those reductions "do not appear to translate into ROI." The survey covered 350 global business executives at organizations with revenue of at least $1 billion that had deployed or piloted AI agents, intelligent automation, or autonomous technology. Gartner projects that more than 40% of agentic AI projects will be cancelled by end of 2027, primarily because of escalating costs, unclear business value, and inadequate risk controls. Only 21% of organizations have a mature governance model for autonomous AI agents.

Poitevin's counter-thesis: "Long term, autonomous business will create more work for humans, not less." Gartner argues that demographic decline and trust-dependent consumer moments will keep human talent central to running and governing AI systems, and that autonomous business will be a net-positive job creator by 2028 to 2029. The practical implication for operators today is not headcount-as-savings but human amplification as the model: the organizations Gartner identifies as on track are those augmenting existing roles rather than eliminating them and building governance infrastructure alongside deployment rather than after the invoice arrives.