U.S. Restores Mythos While Asia Ships Alternatives

Frontier model access is becoming a bargaining chip between governments, and competitors in Asia are treating the interruption as a market opening.

June 28, 2026 | Reading time: 11 minutes | Issue #197

Lead

The U.S. government reversed part of its block on Anthropic's frontier models on June 27, allowing more than 100 U.S. institutions to regain access to Mythos 5, according to Semafor and Axios. Commerce Secretary Howard Lutnick wrote to Anthropic that the company had made "significant progress" in talks and committed to work with the government on "protocols and standards and releases." The letter lifted the Mythos 5 suspension but was silent on Fable 5, the model that had briefly been the most capable consumer-facing system available.

The episode began two weeks earlier, when the Commerce Department's Bureau of Industry and Security issued an export-control directive ordering Anthropic to disable Fable 5 and Mythos 5 for any foreign national, including the company's own foreign staff. Anthropic complied by cutting access for all customers. Legion LegalTech, a San Jose legal-technology company with Canadian developers, filed suit on June 24 calling the harm "immediate, irreparable, and existential," The Next Web reported. The case is notable because it is the first time Washington has used export controls to pull a commercial AI model offline, rather than restricting hardware.

The partial restoration does not end the uncertainty. Fable 5 remains suspended, Axios reported, and the Pentagon and National Security Agency still must approve its return. The criteria for "trusted" status have not been published, and the process for obtaining it is unclear. What is clear is that the U.S. now has a de facto pre-clearance regime for frontier models, run from the White House and Commerce Department rather than Congress. The question is whether this structure can scale as every major lab releases systems with comparable capability scores.

OpenAI's Workforce Now Lives Inside Agents

OpenAI published internal data on June 25 showing that Codex agents account for 99.8% of weekly output tokens inside the company. Every department, including Legal, Finance, and Recruiting, now uses Codex as its primary AI tool. The research paper, "The Shift to Agentic AI: Evidence from Codex," reported that 80.6% of sampled individual users made at least one Codex request estimated to exceed 30 minutes of human work, and 25.6% made one exceeding eight hours. Non-developer usage rose 137x for individuals, 189x for organizations, and 12x inside OpenAI since August 2025.

The numbers suggest a structural shift in how knowledge work consumes compute. ChatGPT interactions are short and cheap; agent runs are long, token-intensive, and recurring. If the pattern holds at customer organizations, inference revenue will depend less on monthly subscriptions and more on continuous execution. The caveat is selection bias: OpenAI employees are the most motivated users of OpenAI products, and the company has not disclosed whether usage is incentivized.

Eastern Front

Asia's AI labs are moving fast while U.S. regulators debate access. Sakana AI, a Tokyo-based startup founded by former Google researchers, launched Fugu on June 22, a multi-agent orchestration model that the company says matches Anthropic's Fable 5 and Mythos Preview on engineering and reasoning benchmarks. The Verge reported that Fugu works by routing tasks across a pool of underlying models, including instances of itself, so that access can shift if any single provider is restricted. Sakana's marketing explicitly frames the product as "frontier capability without the risk of export controls."

China's 360 also unveiled Tulongfeng, an automated vulnerability-discovery tool that the company says can go head-to-head with Mythos, TechCrunch reported on June 27, citing Reuters. Wired reported on June 26 that Chinese users have been circumventing Anthropic's geolocation restrictions to access Claude through increasingly sophisticated workarounds. The parallel developments point to a bifurcation: U.S. labs face voluntary government review, while Asian competitors treat that friction as a feature gap they can fill.

India Lens

OpenAI appointed Prabhjeet Singh, former Uber India and South Asia president, as its first managing director for India, TechCrunch reported on June 26. Singh starts in September and will oversee consumer growth, enterprise adoption, partnerships, regulatory engagement, and operations in what OpenAI calls its second-largest market after the U.S. The hire follows OpenAI's New Delhi office opening in August 2025 and planned offices in Mumbai and Bengaluru.

The appointment signals that India is being treated as an operating market, not a sales territory. OpenAI is investing in policy relationships, data center capacity, and local hiring at a time when the U.S. government is restricting access to the same models. For India, the opportunity is to become a major deployment hub for frontier AI. The risk is that it inherits the access disputes of the U.S.-China rivalry without a seat at the rulemaking table.

Europe

Mistral released OCR 4 on June 23, a compact document-intelligence model with bounding boxes, block classification, inline confidence scores, and support for 170 languages. The model runs in a single container for self-hosted deployments and integrates with Mistral's open-source Search Toolkit. The release is positioned as a sovereignty play: documents stay on the customer's infrastructure, which matters for compliance in the EU and other regulated markets. Reuters reported on June 25 that Anthropic hired Orange's AI chief amid a Europe push, and Bloomberg reported the same day that the U.S. is seeking an AI partnership with the EU on regulation and supply chains. The EU is not matching the U.S. on frontier model capability, but it is building a regulatory and tooling market that U.S. labs may have to enter on European terms.

Compute Watch

Apple is making one of its largest changes to Mac silicon strategy, Bloomberg reported on June 25. The company, currently on its M5 series, plans to debut a base M6 processor for entry-level Macs this year but will skip higher-end M6 chips in favor of an AI-focused M7 Pro, Max, and Ultra lineup in 2027. The move signals that Apple expects on-device AI workloads to drive high-end chip demand, even as it races to catch up to NVIDIA and Qualcomm in AI silicon.

At the same time, Apple is losing talent to AI labs. Bloomberg reported on June 26 that Paul Meade, Apple's vice president in charge of Vision Pro and smart glasses, is leaving for OpenAI's hardware unit. OpenAI has been building a hardware team for months, and the hire suggests consumer AI devices remain part of its roadmap despite the focus on models and agents.

Builder's Corner

Patronus AI raised a $50 million Series B led by Greenfield Partners, TechCrunch reported on June 25, to build simulated digital environments for stress-testing AI agents. The company, founded by former Meta AI researchers, creates replicas of websites and internal systems where agents are evaluated after reinforcement learning training. Patronus says revenue has grown 15-fold over the past year and that virtually every frontier lab is now a customer.

The funding reflects a maturation in the agent stack. Model makers can publish benchmarks, but benchmarks do not prove that an agent will reliably complete a multi-hour real-world task. Simulation-based evaluation is becoming the layer between training and deployment. Netris, meanwhile, raised $15 million from Andreessen Horowitz to automate network setup for AI neoclouds, TechCrunch reported on June 25. The company says its software is already live at more than 35 GPU clusters totaling about a million GPUs.

From the Lab

A paper posted to arXiv on June 25, "When Does Combining Language Models Help? A Co-Failure Ceiling on Routing, Voting, and Mixture-of-Agents Across 67 Frontier Models," puts a statistical bound on multi-model ensembles. The authors show that the accuracy gain of any routing, voting, or cascade system is capped by beta, the rate at which every candidate model is wrong on the same query. Across 67 models, the all-wrong rate on open-ended math was 0.052, about 2.5 times what a standard Gaussian copula would predict. The implication: ensemble gains come from models failing on different questions, not from adding more models that fail the same way. This is relevant to Sakana Fugu's orchestration thesis and to any enterprise building model routers.

The View

Three concurrent developments are reshaping the competitive map. First, the U.S. government is asserting direct control over model release timing and customer eligibility, using informal pressure rather than legislation. Second, the technology is becoming more agentic, more token-intensive, and more deeply embedded in enterprise workflows. Third, global competitors are advancing on separate tracks, capitalized by different funding structures and subject to different release rules.

The tension is between capability and governance. Washington's interventions are designed to reduce catastrophic risk from cyber-capable models, but they also slow the labs that invented them and create a two-tier market in which U.S. citizens get earlier access than foreign nationals. OpenAI's internal Codex data suggests that demand for agentic compute is rising fast; government gating may not be able to hold it back without formal legislation.

The likely outcome is a bifurcated global market. One tier will operate under U.S. government coordination: limited releases, trusted-customer lists, and security reviews. Another tier will operate from China and other jurisdictions with fewer internal constraints. A third tier, including India and Europe, will absorb models from both sides while trying to write local rules. The advantage will go to labs and governments that can make governance itself a competitive feature, not just a cost.

The Miss

California launched an AI-Unemployment Tracker on June 25, linking occupational AI-exposure measures to monthly unemployment insurance claims, Bloomberg reported. The tool was built by Governor Gavin Newsom's office, the state's Employment Development Department, and the California Policy Lab. It is one of the first government dashboards explicitly designed to detect AI-driven labor displacement in real time. The signal it produces will be noisy, and AI exposure indices remain contested, but the institutional response matters. Other states are watching.

Pull Quotes

"I have determined that appropriate safeguards are in place to permit certain trusted partners to access the Claude Mythos 5 Model." — Howard Lutnick, letter to Anthropic, June 27, 2026

"For the average OpenAI worker, Codex usage now accounts for more than 85% of output tokens." — OpenAI, How Agents Are Transforming Work, June 25, 2026

"Access to top models can disappear overnight. Collective intelligence is the practical hedge against this concentration of power." — David Ha, Sakana AI, June 2026

"The harm to Legion is immediate, irreparable, and existential." — Legion LegalTech complaint, June 24, 2026

"Orchestration Models are the next frontier, beyond bigger models." — David Ha, Sakana AI, June 2026

Reads & Links

U.S. releases powerful Anthropic model Mythos to some U.S. companies — Semafor on the conditional restoration of access. https://www.semafor.com/article/06/27/2026/us-releases-powerful-anthropic-model-mythos-to-some-us-companies

Scoop: Powerful Anthropic model, Fable 5, on track to return soon — Axios on the thaw in the administration's standoff with Anthropic. https://www.axios.com/2026/06/27/anthropic-fable-5-return-soon

Statement on the US government directive to suspend access to Fable 5 and Mythos 5 — Anthropic's official account of the export-control order. https://www.anthropic.com/news/fable-mythos-access

How agents are transforming work — OpenAI's internal Codex adoption data and research paper. https://openai.com/index/how-agents-are-transforming-work/

Sakana Fugu: One Model to Command Them All — Sakana AI's launch page and technical report. https://sakana.ai/fugu-release/

Asian AI startups launch Mythos-like models as Anthropic's export ban drags on — TechCrunch on Sakana Fugu and 360's Tulongfeng. https://techcrunch.com/2026/06/27/asian-ai-startups-launch-mythos-like-models-as-anthropics-export-ban-drags-on/

OpenAI poaches Uber India chief to lead its biggest market outside the US — TechCrunch on the India managing director appointment. https://techcrunch.com/2026/06/26/openai-poaches-uber-india-chief-to-lead-its-biggest-market-outside-the-u-s/

Mistral OCR 4 : SOTA OCR for Document Intelligence — Mistral's self-hosted document-intelligence release. https://mistral.ai/news/ocr-4/

Apple to Skip High-End M6 Mac Chips in Favor of AI-Focused M7 Line — Bloomberg on Apple's silicon strategy shift. https://www.bloomberg.com/news/articles/2026-06-25/apple-to-skip-high-end-m6-mac-chips-to-launch-m7-pro-m7-max-m7-ultra-instead

Patronus AI lands $50M to build 'digital worlds' that stress-test AI agents — TechCrunch on the Series B and simulation layer. https://techcrunch.com/2026/06/25/patronus-ai-lands-50m-to-build-digital-worlds-that-stress-test-ai-agents/

When Does Combining Language Models Help? A Co-Failure Ceiling on Routing, Voting, and Mixture-of-Agents Across 67 Frontier Models — arXiv:2606.27288. https://arxiv.org/abs/2606.27288

California State Government Launches AI Job Loss Tracker as Layoff Fears Grow — Bloomberg on the AI-Unemployment Tracker. https://www.bloomberg.com/news/articles/2026-06-25/california-state-government-launches-ai-job-loss-tracker-as-layoff-fears-grow

Frontier model access is becoming a bargaining chip between governments.