4 articles.
Real progress on open conjectures is impressive and mostly irrelevant to the problem set due Thursday. Separating the two is the skill students need right now.
Google Gemini keeps conversations 72 hours when activity tracking is off and 18 months by default, with human reviewers able to see them for up to three years. Anthropic keeps consumer chats 30 days on the backend by default and lets you opt in to model improvement. OpenAI lets any consumer opt out of training while business and API data is not used for training by default. Two of the three describe training as opt-in and one describes it as opt-out.
Transcription APIs are billed in whatever unit suits the vendor: Deepgram and OpenAI quote per minute, AssemblyAI quotes per hour, Google charges per minute with volume tiers that fall by 75%, and Azure quotes per hour with a separate batch tier. Normalised, an hour of audio runs from $0.18 to $2.04, and Google's standard tier is more than four times the price of AssemblyAI's asynchronous model for the same job.
Midjourney's terms say that if you are a company, or an employee of a company, making more than $1,000,000 a year, you must be on Pro or Mega to own your assets, so everyone below that line is generating images it does not hold title to. The same terms commit only to reasonably offering unlimited access while reserving the right to rate limit it. Alongside that: OpenAI bills GPT Image per million tokens and sends you to a calculator for the per-image figure, with Batch at exactly half and cached input only on the Responses API, and Black Forest Labs meters in images per month with a one-domain cap.