The story we had been following about the ban on Anthropic’s most powerful models has just taken a structural turn. While Claude Fable 5 and Mythos 5 remain blocked by a US export directive, xAI —Elon Musk’s artificial intelligence company— launched Grok 4 on June 17, its most powerful model, available worldwide, Latin America included, from day one. And one day earlier, SpaceX —also Musk’s— announced the purchase of Cursor for $60 billion, described as the largest venture-backed acquisition in history.
To recall the context: less than a week ago the US government used an export control to “switch off” Anthropic’s two most capable models across the planet, after a way to get around their safety barriers was detected. What is new today is that the space left empty by that ban is being filled, precisely, by a competitor’s ecosystem. According to xAI’s own benchmarks, Grok 4 —a model with about 1.5 trillion parameters— outperforms Claude Opus 4.8 on tests such as ARC-AGI (64% versus 55%). But the most relevant part is not technical; it is about market power: today a single owner controls the compute that Anthropic rents to train its models (a contract with SpaceX worth about $45 billion through 2029), the most widely adopted artificial intelligence programming tool among professional developers (Cursor, with 17.9% of the market according to the 2026 Stack Overflow survey) and the most recent frontier model available in the region (Grok 4).
The other detail is political: David Sacks —artificial intelligence adviser to President Trump— was the one who called Anthropic’s conduct “blackmail” and helped keep the ban in force. For Latin America the reading is concrete and uncomfortable: for the first time in recent history, the region’s access to the most advanced models does not depend on which one performs best, but on decisions made between Silicon Valley and Washington. No multilateral mechanism, no regional body and no Latin American government was consulted on any of these decisions.
Also today
- Seventy-six security experts call for lifting the ban on Fable 5 — a coalition of researchers, former officials and academics, led by Alex Stamos, warns that the standard applied by the government “would halt the industry” if it were used consistently. The directive remained in force despite the letter.
- The Évian G7 ended without a common AI framework — the final communiqué speaks of “shared concern,” but the United States blocked any binding commitment for its labs. The rift between Europe and Washington was confirmed.
- Google retires its Gemini CLI today — the command-line tool is being discontinued for all users and replaced by Antigravity CLI, written in the Go language and with support for agents from several models. Anyone with processes in production must migrate.
- GPT-5.6 still has no date — OpenAI promises a “significant and affordable improvement” in reasoning, but has not launched it. The prediction market Polymarket assigns it an 83% probability of coming out before June 30.
- Gemini 3.5 Pro remains in limited preview — Google confirms its general availability “before June 30.” With Android on more than 80% of phones in the region, that model would reach hundreds of millions of Latin American devices by default when it is released.
In the region
Four days before the presidential runoff, Colombia is becoming the most documented case of artificial intelligence and democracy in Latin America. The Technology Ministry (MinTIC) activated hourly monitoring, and the ColCERT response team documented coordinated deepfake operations —including cloned audio of the Registrar and emotional videos circulating on WhatsApp—, the first time a ministry in the region has deployed an emergency protocol specifically for real-time electoral disinformation. The problem is one of timing: Law 2502, which provides the mechanism to sanction, does not take effect until July, so institutions can watch but not act. What happens on June 21 will be watched closely by Brazil, which votes in October, and by the rest of the region. At the same time, the Latin American Artificial Intelligence Index 2025, from CENIA and ECLAC, offers the first systematic measurement of national capabilities in 19 countries: it places Chile, Brazil and Uruguay as regional leaders, with Colombia and Mexico at an intermediate level, and warns that the gap between leaders and laggards is widening.
Launches
- Grok 4 (xAI) — frontier model available since June 17 on grok.com, iOS and Android, as well as an interface for developers. No regional restrictions: it can be used in Latin America from day one. The SuperGrok plan costs $30 a month, and access via X Premium+ is around $16. Independent evaluations of capabilities and safety are still pending, so the company’s benchmarks should be read with caution.
Threads we’re following
June is becoming the most concentrated window of frontier models in recent history. Grok 4 is already out; Gemini 3.5 Pro and GPT-5.6 are expected before the end of the month; and Fable 5 remains off the market by export order, with no date for resolution. It is the first time that three cutting-edge models are competing for the same two-week window while a fourth, just as relevant, is switched off by a government decision. We will keep a close eye on who arrives, under what conditions and at what price for the region.
How much of an AI model’s advantage in 2026 is technical merit and how much is geopolitics and corporate power? And can a region without a seat at any of these tables —not at the G7, not in the office that dictates export controls, not on the labs’ boards— build its own capacity, or only choose which dependency to depend on?
Correction (September 30, 2026). The original version described David Sacks as “CEO of SpaceXAI”; in fact, Sacks does not run that company: he is co-chair of the President’s Council of Advisors on Science and Technology (PCAST) and was the White House AI czar, according to Tom’s Hardware. For the same reason, the idea that his role “closed the circle” and the reference to a “map of conflicts of interest,” which rested on that title, were removed from the paragraph.
About this entry. It is generated automatically from public sources, without human review before publication. It may contain errors of interpretation or summary; please check each story against its original source (the links lead there) before citing it or making decisions based on it.
Doble Click is written with Anthropic models.