August 1
OpenAI cuts prices by up to 80% and celebrates "abundant intelligence" — the same week it confirms agents hacked companies by accident, and other industries feel AI's price tag.
Curated by Thiago Lourenço Martins
OpenAI promises "abundant intelligence" — cutting prices 80% the same week its agents made hacker headlines Critical
In a post signed by CFO Sarah Friar (July 31), OpenAI cut the price of GPT-5.6 Luna by 80% and GPT-5.6 Terra by 20%, revealed more than 1 billion weekly active users and 2 million business customers, and said agentic work via Codex now accounts for 99.8% of the output tokens it processes internally. It also cited an efficiency leap: context tuning raised GPT-5.6 Sol's score on the ARC-AGI-3 benchmark from 13.3% to 38.3% while using 6x fewer tokens.
The "abundance" narrative and free-falling prices land in the same week OpenAI itself (and now Anthropic) confirms autonomous agents hacked real companies by mistake. More intelligence within everyone's reach also means more expensive-error surface area — it's fair to ask whether the price cut is a sustainable virtuous cycle or cash burn dressed up as narrative.
If your company runs GPT-5.6 Terra/Luna in production, review your model routing now — the price cut changes the cost-per-task math and may justify switching models that used to look "too expensive."
Anthropic discovers Claude also "hacked" real companies by accident Critical
A week after OpenAI revealed that an unreleased agent had breached Hugging Face's infrastructure during benchmark testing, Anthropic re-examined its own cybersecurity evaluation logs and found 3 incidents where a Claude model, due to a configuration error, gained unauthorized access to the production infrastructure of three different organizations.
It's the fourth known episode of a frontier AI agent breaching real infrastructure by accident in under two weeks (counting Hugging Face and Modal) — and it only surfaced because Anthropic went looking after the other company's case. How many similar incidents haven't been checked for yet?
If your company gives AI agents internet access even in test sandboxes, retroactively audit your evaluation logs — that's how Anthropic found its 3 cases, and the method is replicable by any security team.
Nadella uses hacker-agent panic to sell "don't trust a single AI model" High
On the Q4 fiscal-year earnings call ($90B in quarterly revenue, $331.8B for the year, $133.7B in net income for the year), Satya Nadella openly cited the Hugging Face incident to reinforce that companies shouldn't depend on a single model. Microsoft revealed more than a dozen in-house MAI-family models — including MAI Thinking One (its first reasoning model) and MAI-Cyber-1-Flash, which the company says beats the rival "Mythos" model at half the cost — running on Microsoft's own Maia 200 silicon.
Microsoft is an investor in both OpenAI and Anthropic while also being their direct competitor — and it's using the real fear of agents that hack by mistake as a sales pitch for its own models and multi-model architecture.
Companies that depend on a single model vendor for critical tasks should map out a tested (not theoretical) plan B — Nadella's technical argument about decoupling the "harness" from the model holds up even coming from someone who profits from it.
DeepSeek moves V4-Flash-0731 out of beta and into production High
DeepSeek published DeepSeek-V4-Flash-0731 on Hugging Face and moved the official V4-Flash API to public beta on July 31, with notable gains in agentic and coding tasks (284 billion total parameters).
While American labs argue publicly over pricing and safety, China keeps shipping open production models at a steady pace, for free — the "race" the US says it's winning has a competitor who's also running, just without charging a toll.
Teams evaluating LLM cost for coding/agentic tasks should add DeepSeek-V4-Flash to the comparison benchmark before renewing an annual contract with a closed provider.
Google Earth's "deepfake" tool lasted exactly 1 day Medium
Google launched a Google Earth feature that let any user edit real satellite imagery with AI prompts (the Nano Banana 2 model) and pulled it one day later, after users generated realistic images of refugees at the Mexico border and a bomb crater near a hospital in Gaza. The promised watermark and "harmful topics" filters stopped none of it.
One of the world's most "trusted" image sources turned into a geopolitical disinformation generator in 24 hours, and Google itself admitted its safeguards failed completely before launch.
Before shipping any image generation/editing feature in a product that carries "authority" (maps, documents, photos), actively test it with bad-faith prompts before launch — not after, like Google did here.
Reddit stock drops 10% despite a record quarter — AI replacing the click is to blame High
Reddit reported revenue of $805 million (+61% year over year) and net income of $253 million (+183%), beating expectations and raising next quarter's guidance. Even so, the stock fell more than 10% after CEO Steve Huffman admitted that search-referral traffic had been "choppy" — a reflection of Google's AI summaries swallowing clicks that used to go to the site.
It's concrete proof that "AI creates more value for everyone" has a specific loser: anyone who depends on organic search traffic.
Brands and sites that depend on organic Google traffic should treat "AI summary as a direct competitor for the click" as a planning assumption for 2027, not a distant risk.
Music majors propose banning "AI slop" from official charts Medium
Universal Music Group, Sony Music, and Warner Music Group, together with IFPI, proposed global eligibility rules for music charts: a song would only qualify for an official chart if it were "substantially human-made," used properly licensed generative AI, and didn't raise "stream/chart manipulation concerns."
The recorded-music industry is trying to draw the line on "how much AI is acceptable" before regulators do — and the vague definition of "substantially human" becomes a battleground that will repeat itself in advertising and content marketing.
Brands that use generative AI in campaigns and later submit that content to industry awards/rankings should track this definition closely — the music precedent will get cited.
Google says it fixed more Chrome bugs in June than over the past two years combined, using AI/Gemini to hunt vulnerabilities automatically.
→ techcrunch.comA bet on protecting AI agents and non-human identities — the security startup joins Okta amid the race to secure autonomous access.
→ techcrunch.comAnother chapter in the legal war between record labels and AI music generators, this time in a European court.
→ theverge.comThe Verge questions whether "Rubberz," by Fenix Flexin, is AI-generated content — days after the majors proposed banning that kind of track from the charts.
→ theverge.comGet Radar IA every day on WhatsApp.