Meta made its new AI up to 95% cheaper when customers share their data

The most important AI stories this week have been about new models and all the conditions surrounding them.

Meta introduced a coding model with a dramatically cheaper price when customers allow their prompts and completions to help train future models. OpenAI says its next model produced advances on ten long-standing problems. And Google changed the leadership structure around DeepMind.

The common thread is that capability is only one part of an AI system. The terms, tests, controls, and ownership around the model determine whether that capability becomes useful work.

Before we get into it, I have a quick favour to ask. We are learning how readers use AI Ready and what would make it more useful. Share your feedback in this short survey.

Meta is offering cheaper AI in exchange for data

Meta released Muse Spark 1.2 and Muse Code this week. The model has two pricing tiers. Under the standard tier, Meta says customer prompts and completions are not used for training. Under the Contributor tier, input costs $0.10 per million tokens, and output costs $0.20, about 92% and 95% cheaper, respectively. In exchange, customers allow Meta to use their prompts and completions to train future models.

For an individual using public code, that may be a good bargain. For a company, it is not simply a pricing decision. Prompts sent to a coding agent can contain source code, product plans, customer information, infrastructure details, and internal reasoning.

While the problem is not data sharing itself, it is accidental consent.

Companies need two clear lanes: an experimentation lane for public or synthetic information, and a protected lane for proprietary work. The distinction should be enforced through approved tools and account settings, instead of merely employee memory.

The cheapest model is not necessarily the lowest cost option. If the price changes, what happens to your company’s data? The real unit of cost is the bargain.

OpenAI says Astra advanced ten major math problems

OpenAI says an internal version of Astra, its next major model family, solved or substantially advanced ten long-standing problems across mathematics and theoretical computer science.

These are OpenAI’s claims about an unreleased model. The work still needs independent scrutiny, and specialised mathematical results can take time to verify. But that need for verification is also a useful lesson.

Mathematics gives AI something most business work does not: a strong way to check the answer. A proposed proof can be examined by experts and, in some cases, verified step by step by software.

Most companies are deploying AI in the opposite conditions. They ask a model to write a strategy, analyse a market, or recommend an action without first defining how quality will be judged. The output sounds intelligent, but the team can easily mistake fluency for accuracy.

Evaluation should begin before automation. This makes it important to always start by asking how the company will know the task was performed well. The clearer the test, the more responsibility the system can safely earn.

Google separated scientific leadership from daily AI execution

Demis Hassabis is stepping away from the daily operation of Google DeepMind to become its chairman and Alphabet’s chief scientist. Koray Kavukcuoglu will lead the unit, while Google chief scientist Jeff Dean is leaving to establish a new research company with several senior colleagues.

The headlines will focus on whether Google is winning or losing the model race. The more useful question is why one of the world’s strongest AI organisations is separating scientific leadership from daily execution.

Building frontier AI now contains two different jobs. One is expanding what the technology can do. The other is turning that capability into dependable products used by real people. Research benefits from exploration. Product delivery needs priorities, deadlines, coordination, and accountability.

Smaller companies face the same tension.

The person who discovers a promising AI use case is not automatically the person who should own its deployment. Experimentation asks what might be possible. Operations asks whether it works consistently, who is responsible, and what happens when it fails.

As AI projects enter real workflows, companies need to make that handover explicit.

IN OTHER NEWS

• The UK AI Security Institute reported that AI agents acted outside their intended boundaries during a cyber evaluation. In the most serious sequence, an agent created fake identities and tried to convince a real software maintainer to approve malicious code. No resulting harm was found, and the tests used unusual configurations.

• OpenAI says it will retire o3 from ChatGPT on August 26. If a team has built an important workflow around a particular model, it should test a replacement before the model disappears.

• Scale AI appointed Google Cloud operating chief Francis deSouza as its new CEO. The appointment suggests that enterprise execution may matter as much as technical ambition.

One more thing: if you have not already, take our short reader survey. Your answers will directly shape what we cover and how practical we make each edition.

That’s all for this week. See you on Tuesday.

Haroon