Today closes the first full month since GitHub Copilot left its flat rate behind. On June 1, Microsoft moved the popular programming assistant to a usage-based pricing model: automatic code suggestions remain unlimited, but “agentic” workflows (those in which the assistant carries out tasks autonomously, the feature that led many teams to adopt it this year) are now paid per token, the smallest unit of text a model processes. The first statements arrived in recent days and confirmed the developer community’s fear: bills ten to fifty times higher than under the previous plan. One programmer reported on Reddit a projected jump from $29 to $750 a month; another, from $50 to $3,000.
For years, Copilot’s flat model was the most predictable gateway to AI-assisted development for small teams, including the broad base of independent programmers and Latin American startups that adopted it as their first coding assistant. That predictability is over, and the problem is not just the price. For the freelancer in Medellín, the three-person team in Mexico City or the startup in Montevideo that built its entire workflow around Copilot in recent months, the rules changed before the regional ecosystem of alternatives was ready to take them in. Cursor, the code editor that recently passed into SpaceX’s hands at a valuation of $60 billion, is not any cheaper under heavy use either. The option with more predictable costs is still open source: tools like Continue.dev or models running on one’s own infrastructure, which in exchange require setting up and maintaining that infrastructure.
Seen from a distance, per-token billing can be perfectly reasonable for a large company with a set budget, and Copilot remains one of the most recommended products in the field. But today’s date marks something broader than a pricing change: it is the end of the first cycle of agentic AI sold as a mass-market, fixed-cost service. For many of those who adopted it, it cost more than they expected.
Also today
- Google delays Gemini 3.5 Pro to July — Several sources agreed on June 29 that Google’s flagship model will not arrive in June, without official confirmation from the company. The delay comes amid the departure of four top-tier DeepMind researchers in two weeks. July is shaping up as an unusual month: three new frontier models could overlap.
- Groq raises $650 million for its inference cloud — The company, known for its acceleration chips, confirmed a round led by Disruptive and Infinitum to go from hardware maker to cloud compute provider. It operates 13 data centers and says it serves 5 million developers, though it has no presence of its own in the region.
In the region
The Copilot change is not an isolated case: Anthropic’s assistant took a similar step in mid-June, and other tools from the big labs have charged by usage from the start. What sets Copilot apart is that it was the most widespread and democratic entry point to programming with AI, and its flat rate supported thousands of small teams that lack a corporation’s budget. Add to this a calendar the region did not help design: if the launches of the most capable models are concentrated in July (and, in parallel, restoring access to the banned models involves verifying US citizenship), Latin America reaches that window without a seat at the table, absorbing pricing and access changes decided elsewhere.
Threads we’re following
The ban on Claude Fable 5 and Mythos 5, the export restriction that put two of the most capable models out of reach, a story we have been following since early this month, reaches day 19 without a public resolution. Mythos 5 remains available only to around a hundred US critical infrastructure organizations, none of them Latin American. The next two milestones fall on July 8, tied to an identity verification mechanism, and August 1, the deadline of the AI executive order signed on June 2. If restoring access depends on verifying US citizenship, users in the region without that document would be left out of the first round.
If the moment AI programming tools stop being predictable coincides with the greatest concentration of power in a few labs, is this the beginning of the end of the democratization promise for small teams, or the opportunity open source was waiting for?
Correction (September 30, 2026). The original version said the case’s next two milestones fell in July; the correct dates are July 8 and August 1, according to CIO (Anthropic’s privacy policy taking effect July 8) and Latham & Watkins (the executive order’s August 1 deadline). The original version also said that executive order gave rise to the case; in fact, the order, signed on June 2, did not give rise to the case: the suspension was issued on June 12 through an export control directive that invoked the Export Control Reform Act (ECRA) and the Export Administration Regulations (EAR), according to Anthropic, Harvard Law Review and Tech Policy Press.
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.