Why these five developments made the list
AI news moves quickly, but speed makes it easy to mix announcements, rumours and predictions. This roundup includes only events announced or put into effect during August 2026 and supported by an official source.
Official sources still need context. A company can accurately report what it released while presenting the result in the most favourable way. For that reason, each section separates the confirmed event from the claims and conclusions that still need independent testing.
Unconfirmed model names, September predictions, an OpenAI stock-market filing reported before August and July’s AI-controlled F-16 announcement. They may be interesting, but they are not five confirmed August developments.
OpenAI reported ten advances in mathematics
On 1 August, OpenAI published ten results produced with an internal version of Astra, a model that was not publicly available at the time. The problems cover areas including geometry, coding theory, cryptography and quantum complexity.
The careful wording matters. OpenAI said each result either resolves or makes substantial progress on a long-standing open problem. That is not the same as saying that Astra completely solved ten universally accepted unsolved problems.
What is confirmed?
- OpenAI released manuscripts, reasoning walkthroughs and machine-checkable Lean certificates.
- Humans used the same model to prepare the arguments as manuscripts.
- OpenAI estimated that the solution-generating tokens would cost roughly $2,000 at its Sol API rates.
What does that not prove?
A Lean certificate can check whether a formal proof follows the rules encoded in the system. It does not by itself decide whether a result is important, original or presented in the best mathematical context. OpenAI also said it hoped the mathematical community would examine and place the work in context.
This is a concrete example of AI moving beyond summarising existing material and contributing to research where parts of the result can be checked formally. It is not evidence that one model is correct on ordinary questions that do not have machine-checkable answers.
Palantir reported 93% year-over-year revenue growth
On 3 August, Palantir reported second-quarter revenue of $1.935 billion. That was 93% higher than the same quarter a year earlier and 19% higher than the previous quarter. Its United States commercial revenue rose 149% year over year to $764 million.
Year over year compares a period with the same period one year earlier. It is useful because it reduces seasonal distortion. It does not mean revenue rose 93% every quarter.
What is confirmed?
The figures come from Palantir’s quarterly investor materials. They show rapid growth at one company that sells data and AI software to commercial and government customers.
What remains limited?
Palantir’s results do not prove that all AI companies are profitable, that AI caused every dollar of growth or that the same rate can continue. The investor release is also company-authored, even though quarterly financial reporting is more structured than a promotional product post.
The figures show that some organisations are spending substantial money on deployed AI and data systems—not only experimenting with free chatbots. They should not be used as a shortcut for judging the whole industry.
Important EU transparency duties began applying
Article 50 of the European Union’s AI Act began applying on 2 August. The rules cover particular providers and deployers of AI systems; they do not simply require every piece of AI-assisted content to carry the same visible label.
What do the rules cover?
- People must be informed when they are directly interacting with certain AI systems rather than a human.
- Providers must add machine-readable marks that help detection tools identify generated or manipulated content.
- Deployers must disclose exposure to deepfakes and certain AI-generated text about matters of public interest.
- Separate duties apply to emotion-recognition and biometric-categorisation systems.
The European Commission’s guidance includes definitions, examples and exceptions. Systems placed on the market before 2 August also receive a limited grace period for the machine-readable marking obligation until 2 December 2026.
The exact duty depends on the system, content, role and use. Standard editing and some creative or satirical work may be treated differently.
People in the EU should increasingly see clearer disclosure when a service is an AI system or when particular public-interest content has been generated or manipulated. A label still does not tell you whether the content is true.
Google changed who runs DeepMind day to day
On 5 August, Google announced new roles across its AI leadership. Demis Hassabis moved from day-to-day operational responsibility to become Chair of Google DeepMind and Chief Scientist of Alphabet. He also continues to lead Isomorphic Labs.
Koray Kavukcuoglu became Senior Vice President of Google DeepMind, reporting to Alphabet CEO Sundar Pichai. Google said he would oversee Gemini model development, frontier AI research, the Gemini app and developer teams.
What did not happen?
Hassabis did not leave Google DeepMind, and the organisation did not close. His own message described the new position as a strategic role focused on broader AI and scientific questions while Kavukcuoglu takes operational responsibility.
Leadership changes can influence research priorities and product decisions, but an organisation chart does not predict whether the next Gemini release will be better. The useful fact is the change in responsibilities—not speculation about private motives.
Google released Gemini 3.7 Flash
Google introduced Gemini 3.7 Flash on 13 August. It described the model as a faster workhorse for coding, agents, document work and web development, with introductory API pricing of $0.75 per million input tokens and $3.75 per million output tokens through the end of 2026.
Who could use it at launch?
- Developers through the Gemini API, Google AI Studio, Android Studio and Google Antigravity.
- Enterprise customers through Google’s enterprise products.
- Google AI Pro and Ultra subscribers through Gemini Spark in supported countries.
Google published benchmark improvements over Gemini 3.6 Flash. Those results are useful evidence about Google’s testing, but they remain selected company-reported benchmarks. Performance on your own files and tasks may differ.
Model numbers now change quickly. You do not need to switch every time a decimal changes. Check whether the model is available in the product you use and whether it improves your real task before changing plans or paying more.
Confirmed announcements are not finished stories
Mathematical acceptance
Researchers still need time to study the significance, novelty and wider context of OpenAI’s results.
Business durability
One strong Palantir quarter cannot show whether its growth rate—or wider AI spending—will continue.
Rule enforcement
Real cases will clarify how regulators apply the EU transparency rules across different products and content.
Product performance
Independent use will show where Gemini 3.7 Flash improves daily work and where Google’s benchmarks do not transfer.
The practical lesson is simple: read the date, open the primary source and look for the boundary between “the organisation announced this” and “independent evidence proves this”. Our five-step fact-checking guide gives you a repeatable method.