What was a leak last week is today a product with a name and a price. OpenAI unveiled GPT-5.6, split into three models (Sol, Terra and Luna), and confirmed what had been anticipated: it went straight to an access list of about twenty organizations approved by the White House, with no availability for the general public or for Latin America (OpenAI). This completes a pattern that began two days ago: on June 26, the most capable models of the two most influential frontier labs, OpenAI’s GPT-5.6 and Anthropic’s Mythos 5, were placed on the same day under access lists approved by the US government.
The detail that did not exist before is pricing, and that is where the regional angle appears. Sol is the flagship model ($5/$30 per million input/output tokens) and sets a new record on a standard benchmark of programming tasks; Terra comes in at half the price of the previous generation; and Luna, at $1/$6 per million, falls in a range competitive with regional models and with several open-source alternatives. When access is expanded, Luna will probably be Latin American developers’ gateway to this family. The problem is the “when”: OpenAI only says “in the coming weeks,” with no date, no criteria and no mechanism for consulting the region. The difference from the Fable 5 ban, which was reactive and followed a safety failure, is that here the control is preventive: the list of who gets in was defined before the launch.
Meanwhile, the map of what can actually be used in the region is rearranging itself. This week, independent evaluations documented that GLM-5.2, an open-weights model from China’s Zhipu AI (under an MIT license), matched Mythos 5 on safety tests, the first open model to reach that threshold, and it is available without any government list (explainX). This is not a deliberate choice by the region’s developers but the result of decisions made in Washington: the open model filling the safety gap is Chinese, and the high-end option accessible without an approved company ended up being Grok.
Also today
- Grok 4.3 comes to SuperGrok and X Premium+ with no business requirements — It is the first mass access to the Grok 4 line in the region: with Fable 5 blocked, it is the model with the best score on the ARC-AGI reasoning test (64%) available to individual paying users, with no need to be a “trusted partner.”
- Baseten closes a $1.5 billion round at a $13 billion valuation — The largest inference infrastructure round to date; the company processes a billion calls a day and grew its revenue twentyfold in a year. No presence of its own in LatAm.
- Japan approves a $2.3 trillion AI and semiconductor plan through 2040 — The largest sovereign AI plan in history. For scale: Mexico’s Coatlicue plan (6 billion pesos over two years) amounts to about 0.1% of what Japan plans to invest over the same horizon.
- OpenCode surpasses 160,000 stars on GitHub and 7.5 million monthly users — Its adoption curve coincides with the start of the Fable 5 blockade: it is the first quantified sign of developers migrating to open-source tools during the ban.
- Claude Code ships its first major update since the ban — It lets administrators limit which model versions their team uses and fixes a bug in structured output generation. With Fable 5 blocked, it runs on Opus 4.8 and Sonnet 4.6 for all users in the region.
In the region
June 26 produced two divergent responses to the same problem. That day, Latin America’s technology ministers adopted, in Santo Domingo, the Declaration of the UNESCO summit on AI ethics, their first attempt to set a common position. A few hours later, OpenAI launched its most capable model under a list controlled by the White House. The Declaration contains no mechanism that has a say in who gets onto those lists, and its roadmap has a twelve-month horizon. The concrete result, today, is that the model with safety capabilities equivalent to Mythos 5 and available without restrictions in the region is open-weights and of Chinese origin. Access is not being decided at the table where the region agreed on its principles.
Launches
- GPT-5.6: Sol, Terra and Luna (OpenAI) — A family of three models with tiered pricing ($5/$30, $2.5/$15 and $1/$6 per million tokens). Currently inaccessible from the region because of the government list. The three-tier scheme within a single family is a structural novelty worth comparing, when possible, against regional and open-source alternatives.
Threads we’re following
The blockade of Fable 5, the Anthropic model banned early this month over a safety risk, reaches seventeen days without resolution for the general public. This week’s news comes not from the ban itself but from what surrounds it: an open-source model has already matched the blocked system on safety, and open tools are growing precisely in the gap it left. The next announced milestone in the process is July 8.
If access to the most capable models is increasingly defined by lists drawn up in Washington, what concrete mechanism (prior consultation, a collective regional voice, or formal representation) would need to exist so that the next “June 26” is not decided, once again, without the region in the room?
Correction (September 30, 2026). The original version said the Coatlicue plan cost $6 billion over five years and amounted to 0.9% of what Japan plans to invest over the same horizon; the correct figure is 6 billion pesos (about $326 million) over two years, according to El Financiero, which amounts to about 0.1% of Japan’s investment over that period.
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.