July 13
OpenAI cuts prices to win the cheap-model war — the same day Apple sues it for stealing hardware secrets.
Curated by Thiago Lourenço Martins
Price war: OpenAI, Meta, and xAI slash the cost of their AI models Critical
In recent days the race for cheaper AI turned into an open fight: OpenAI made GPT-5.6 more token-efficient, charging less per task; xAI launched Grok 4.5 with twice the efficiency of the previous generation; and Meta launched Muse Spark 1.1, a low-cost agentic model via its new Meta Model API. In the same move, OpenRouter — a service that routes calls across models from dozens of providers — revealed it had raised more than US$ 100 million in May, driven by growing demand for more affordable options.
For the first time, the three biggest general-purpose AI giants are openly competing on cost, not just benchmarks. That changes the math for any company running AI in production — the same result can come out much cheaper just by switching models or providers.
If your company uses AI at volume (support, content generation, automation), it's worth reviewing this month's bill now — comparing cost per task between your current models and the new versions can cut spend without losing quality. Routing tools like OpenRouter help automate that choice.
China tightens the squeeze: GLM-5.2 and Qwen3.6 arrive beating Western models — and companies are already switching to save High
Zhipu AI (Z.ai) launched GLM-5.2, an open model that, according to independent benchmarks, beats Google's best models at reasoning — announced the same day the company closed a US$ 4 billion stock sale. Almost simultaneously, Alibaba launched Qwen3.6, with the Plus version offering 1 million tokens of context and Max-Preview leading agentic coding benchmarks like SWE-bench Pro. A Financial Times report, which went viral on Reddit, shows Western companies swapping expensive closed models for these open Chinese ones to cut costs.
The quality gap between Chinese open models and American closed models has shrunk too fast to ignore — and that puts direct pressure on the price you pay for AI today, no matter the provider.
Before renewing a contract with a closed provider, run a test comparing GLM-5.2 or Qwen3.6 on your most expensive task (support, code generation, document summarization). The cost difference might even justify hosting the model yourself.
Apple sues OpenAI for stealing hardware secrets — and targets OpenAI's own hardware chief Critical
Apple filed a 41-page lawsuit against OpenAI, accusing three former employees — including Tang Tan, now OpenAI's head of hardware — of leaking confidential Apple hardware design documents to help develop OpenAI's own AI device.
It's the first heavyweight lawsuit from a big tech company against an AI lab over industrial hardware espionage — a sign that the race for the "AI device" (the possible successor to the smartphone) is already a billion-dollar fight, with real legal risk for anyone hiring competitors' former employees directly into sensitive areas.
If your company hires people from direct competitors for R&D or product roles, reinforce your confidential-information non-use clauses now and document the provenance of technical knowledge brought in — the Apple vs. OpenAI precedent is about to become a reference point for employment contracts across the industry.
Meta pulls Instagram AI feature days after launch under privacy pressure High
Meta pulled "Muse Image," a feature that generated AI images from public Instagram posts, less than a week after launch — following a strong backlash from users concerned about personal content being used without explicit consent.
It shows the line between "using a user's public data" and "violating user trust" has gotten thinner — and that even a giant like Meta backs off quickly when the backlash is strong enough. It's a warning for any company launching a generative AI product built on user data.
Before launching any AI feature that uses user content (even public content), test the reaction with a small group and make clear, in plain language, exactly what's being used and how to opt out. A public reversal costs reputation — prevention is cheaper.
Helsing raises US$ 1.8 billion — Europe's largest defense round targets Anduril rival High
German defense startup Helsing raised US$ 1.8 billion, reaching a valuation of US$ 18 billion — the largest investment round for a European defense startup to date. The company builds software- and AI-driven defense systems and is seen as the main European rival to America's Anduril.
It confirms that investment in "defense AI" isn't just an American phenomenon — Europe is building its own stack of AI-driven military technology suppliers, at a moment of rising geopolitical tension and pressure for the continent's strategic autonomy.
Anyone working in venture capital, defense, or critical infrastructure should watch this sector closely — the valuation bar for "defense AI" just jumped significantly, which tends to pull up rounds for competitors and supply-chain vendors.
Delaware wants to create the "LLC for AI agents" — and the US Senate proposes the 1st federal law to regulate autonomous agents High
Delaware's Secretary of State, in partnership with legal-compliance startup Norm Ai, proposed a new legal structure — dubbed AIC (AI Agent Company) — to grant legal personhood to AI agents that already conduct business autonomously, inspired by the historic creation of the LLC. The same day, Senator Mark Warner introduced the AI AGENT Act, the first US federal bill aimed specifically at regulating the powers of autonomous AI agents.
Today an AI agent that signs a contract, makes a payment, or negotiates on a company's behalf operates in a legal gray zone. If these proposals move forward, we'll have, for the first time, a clear "legal status" for who — or what — can act autonomously on a business's behalf.
If your company already uses or plans to use AI agents for tasks with real decision-making power (purchasing, negotiation, support with authority to resolve disputes), start documenting those operations now — whatever legal framework comes will demand traceability from day one.
More than 200 experts, including 15 Nobel laureates, call for urgent action on AI's economic impact High
More than 200 researchers and economists — including 15 Nobel Prize winners and scientists from within OpenAI, Anthropic, and Google themselves — signed a public appeal urging governments to act urgently on AI's economic effects on the labor market and income distribution.
When scientists from the very labs building AI sign the warning alongside outside economists, the signal is stronger than the usual "existential risk hype" — it's people with direct access to the technology saying the economic impact is already a problem now, not a distant future one.
HR and strategic planning leaders should treat this kind of signal as a trigger to get ahead of workforce reskilling scenarios — waiting for public policy to move first tends to leave your company playing catch-up later.
For the second time in a week, Anthropic extended free access to Fable 5 across all paid plans, with Claude Code limits 50% higher — a direct counterpunch to the launch of GPT-5.6 Sol.
→ forbes.comThe traffic app now accepts conversational voice incident reports, natural-language destination search, a "less chatty" mode, and a specific mode for motorcycle riders.
→ theverge.comThe world's largest contract chipmaker beat market expectations, with demand for AI chips as the main driver behind the results.
→ reuters.comCanada's federal banking supervisory body warned major financial institutions about advanced AI risks, using Anthropic's model as a concrete reference point for capability.
→ reuters.comIn a post on X, Elon Musk publicly acknowledged he now considers Anthropic's models the most capable on the market today — a notable turnaround from someone who was once a fierce critic of the company.
→ finance.yahoo.comGet Radar IA every day on WhatsApp.