AI Model Price War | AI Field Notes #67

AnalysisVoice assistants have always taken turns: you talk, then it answers. GPT-Live, which OpenAI shipped the week of July 13, drops the turn-taking. It listens and speaks at the same time, a design called full-duplex, and can translate live, search the web mid-sentence, or hand a task to another agent while the conversation keeps going. The interruption gap is what made older voice bots feel like walkie-talkies, and closing it moves the product from party trick toward something you would leave running in a meeting. The open question is what an always-listening model does to anyone whose job is talking to customers for a living.

AnalysisFive enterprise software giants, Google, Microsoft, Salesforce, Snowflake, and ServiceNow, agreed on July 13 to back a shared protocol for how AI agents plug into business data. The target is clear: Anthropic's Model Context Protocol and OpenAI's connectors, which have quietly become the default way agents reach tools and databases. Whoever controls that layer sets the terms for everyone building on top. For a developer wiring an agent into company systems this week, it means one more competing standard to support before any of them has actually won.

AnalysisMore than 400 former Apple employees have left for OpenAI, and on July 11 Apple decided that was theft rather than a hot job market. The suit accuses OpenAI of using Apple trade secrets that walked out the door inside departing engineers' heads, the kind of case a company files when the poaching starts to hurt. Within days Elon Musk and Sam Altman were trading insults on X, turning a legal filing into a weekend show. Under the noise sits a real shift: the most valuable thing a frontier lab can take now is the people who know how the work was done.

AnalysisTrained partly on how developers actually use Cursor, the AI code editor SpaceXAI now owns, Grok 4.5 arrived on July 8 as a coding specialist at $2 per million input tokens and $6 output. It scores 83.3% on Terminal-Bench 2.1, a test of whether a model can finish real command-line tasks, and runs on a 1.5-trillion-parameter mixture-of-experts design, which activates only part of itself per request to hold costs down. The tell is the training data. SpaceXAI bought Cursor, watched millions of coding sessions, and turned them into a cheaper model aimed straight back at those same developers.

AnalysisChinese-built models now handle between 30% and 46% of the API traffic on US developer platforms, according to a CNBC investigation published July 7. On OpenRouter, a service that sends each request to whichever model a developer chooses, DeepSeek's V4 Flash alone takes 17.6% of all tokens and costs $0.14 per million input, roughly 36 times cheaper than OpenAI's GPT-5.5 at $5. For American engineers, the driver is cost, plain and simple. The complication landed fast: China's commerce ministry is reportedly weighing limits on overseas access to its best models, which would strand anyone who built on cheap and assumed cheap would stay.

AnalysisThe part of Europe's AI law with real teeth switches on August 2, and a lot of companies shipping AI into the EU are not ready for it. From that date the rules for general-purpose models and the transparency duties both go live, backed by an AI Office that can levy fines. Brussels paired the deadline this month with a cybersecurity action plan for advanced models and a push to build its own model-evaluation capacity by 2027. Smaller firms caught a break, since the simplified track now stretches to companies with up to 750 staff. Everyone else has about three weeks to learn what compliance actually requires.