AI Model Cost Gap | AI Field Notes #68

AnalysisTwo frontier models did the same coding task last week and the bills came back nearly five times apart: $2.49 on xAI's Grok 4.5, $11.80 on Anthropic's Claude Fable 5. The gap traces to tokens, the chunks of text a model reads and writes and gets charged for. Grok finished in 15,954 output tokens where Fable burned 67,020, roughly 4.2 times more. One developer logged $16 on OpenAI's Sol against $63 on Fable for the same job. Capability has converged at the top of the market; what still splits the field is how much thinking each model bills you for.

AnalysisVoice assistants have always taken turns: you talk, it waits, it answers. OpenAI's new GPT-Live drops the turns. Built on a full-duplex design, meaning it can listen and speak at once, it handles interruptions, real-time translation, and live web search mid-conversation, and it can hand tasks to other agents while you keep talking. The obvious target is the phone call and the live meeting, where half a second of latency breaks the illusion of talking to a person. OpenAI shipped it alongside the GPT-5.6 family in early July. Whether it holds up outside a staged demo is the open question.

AnalysisAnthropic keeps giving its top model away, and the reason is sitting one launch over. It extended free access to Claude Fable 5, its most capable flagship, through July 19, the third extension in five weeks since the original June 22 cutoff, handing Pro, Max, Team, and Enterprise users half their weekly limits at no charge. The trigger each time has been an OpenAI move; this round follows Sol, the new GPT-5.6 flagship. Giving away your best model for five straight weeks is a fight for default habit, waged while switching costs are still low enough for users to walk.

AnalysisStarting today, China draws a line between the AI that does your work and the AI that keeps you company, and only the second kind is banned. Beijing's Interim Measures for AI Anthropomorphic Interactive Services took effect July 15, and ahead of it ByteDance switched off the companion agents inside Doubao, the country's most popular chatbot, while Alibaba pulled the same features from Qwen. Doubao users keep read-only access to their old agent histories until October 15, after which the data goes. The rules, co-issued in April by China's cyberspace regulator and four other agencies, target emotional simulation and leave task automation alone.

AnalysisOver four hundred former Apple employees now work at OpenAI, and Apple's new lawsuit says some of them did not leave empty-handed. Filed July 10 in federal court, the complaint accuses OpenAI of a coordinated push to lift trade secrets tied to unreleased AI hardware: coaching departing staff to dodge Apple's security checks and, in one case, keeping a work laptop loaded with confidential technical documents. Apple names OpenAI's chief hardware officer, Tang Tan, and points at the Jony Ive design venture OpenAI absorbed. OpenAI calls the claims baseless. The suit reads as Apple conceding it lost the talent race and reaching for the courts to slow the winner.

AnalysisThe assumption that the best image and video AI ships from San Francisco is quietly breaking. Kling, the Chinese video generator, raised $2.8 billion in a round backed by Alibaba and Tencent to expand its operations, one of the largest checks written into any AI startup this year. Days earlier ByteDance released Seedream, its latest image model, matching Western tools on quality while charging less to run. Text generation still favors US labs, but pixels are the one domain where Chinese models now set the pace. For anyone generating art, ads, or product shots, the cheapest frontier option increasingly ships from China.