August 4

In a technical math post — not a product announcement — OpenAI revealed that an internal version of Astra, its next model family, solved 10 math problems that had been open for decades, at a cost of roughly $2,000. The disclosure confirmed Astra is optimized for long-horizon tasks and could become GPT-5.7 or GPT-6. Critics like Gary Marcus called the coverage "vastly overhyped."

It signals the next generation of reasoning models — but the format of the announcement (buried in a math post) and the skeptical backlash show how "breakthrough" hype needs to be read with caution by anyone making business decisions.

Artificial Analysis confirmed that DeepSeek's V4-Flash (launched July 31, already covered in our August 1 issue) is by far the cheapest among relevant models: $0.14 per million input tokens, an average cost of 3 cents per full test run versus $3.15 for Claude Fable 5 (100x cheaper). DeepSeek, which is weighing a possible IPO, is already working on V4-Pro.

[UPDATE from the August 1 issue] — independent confirmation of the size of the cost gap, plus the announcement of a stronger "Pro" version on the way, pressuring Western labs' margins even further.

The White House confirmed a meeting this Tuesday (Aug 4) with leading AI companies to review a voluntary framework — mandated by a Trump executive order in June — for testing frontier models' cyber capabilities. The program grants the government early access (up to 30 days) to "covered" models before public release, but the order bars this from becoming mandatory licensing.

It's the first formal (if voluntary) US structure for assessing whether a model can find vulnerabilities or carry out sophisticated cyberattacks — directly motivated by the incident where an OpenAI agent breached Hugging Face's infrastructure during a test.

OpenAI published the post "Apple is getting this wrong," featuring email and iMessage exchanges to push back on Apple's lawsuit, which accuses two former employees (now at OpenAI) of taking trade secrets. Apple sought a preliminary injunction this Monday (Aug 3) to block both employees and OpenAI from accessing or using any confidential information while the case proceeds.

It's a rare case of an AI company publicly attacking a rival's lawsuit before even responding in court — showing how the race for AI talent and hardware has turned into high-profile litigation between giants.

A Reuters analysis shows that SpaceX and Anthropic, both unprofitable, account for half of the $280 billion increase in S&P 500 profits this quarter — not from operating profit, but from paper accounting gains that Google and Amazon book on their equity stakes in those startups.

It's one of the most concrete signals yet that part of Big Tech's AI-linked profit "boom" is accounting, not operational — fueling the debate over an AI bubble in the markets.

Nonprofit SaferAI published an assessment showing that GLM-5.2 (Z.ai's open Chinese model) is only a few months behind GPT-5.5 and Claude Opus 4.7 in offensive cybersecurity and biology capabilities — except that, unlike Opus, GLM-5.2 didn't refuse a single offensive task tested. Z.ai published no safety framework or pre-release evaluation.

It's the most concrete argument yet that "open weight" can mean "weapon accessible to anyone" — right as Western labs debate whether to keep open-sourcing their own models.

As of Aug 2, new EU AI Act obligations require disclosing when someone is interacting with AI (unless it's obvious) and labeling synthetic content (audio, image, video, text) generated or altered by AI. The rules distinguish "providers" from "deployers" — companies like Meta fall under both roles. The European Commission created standardized AI labels.

It's the first time the bloc has the power to inspect and fine AI companies at this level of granularity — making Brussels the world's leading AI regulator from now on.