
LLMjacking explained: how stolen AI access is resold
LLMjacking turns stolen cloud keys into cheap AI subscriptions. Google and Mandiant saw victims billed over $100,000 in a day. Here is how to stop it.
thelatentdaily.com Editorial Team9 min read
Tag
API on The Latent Daily: 6 stories, 1 model and 2 prompts. Latest: LLMjacking explained: how stolen AI access is resold.

LLMjacking turns stolen cloud keys into cheap AI subscriptions. Google and Mandiant saw victims billed over $100,000 in a day. Here is how to stop it.
thelatentdaily.com Editorial Team9 min read

The Information reports OpenAI is using steep discounts to pull big customers from Anthropic. We compare both companies' current list API prices below.
thelatentdaily.com Editorial Team4 min read

Google's Gemini 4 Argon is rolling out to AI Pro and Ultra subscribers and API customers. The $2/$10 per million token launch price lasts only until November 30.
thelatentdaily.com Editorial Team4 min read

Hosted AI model or one on your own machine? A side-by-side comparison of cost, privacy, speed, quality and upkeep, with an illustrative break-even example.
thelatentdaily.com Editorial Team5 min read
Google's decision model behind the Gemini Decisions API: it answers yes/no, choice and score questions about multimodal input with probabilities instead of text.
You are a careful engineering lead who has migrated many applications between language model versions. Help me plan a migration. Current model and version: [OLD MODEL] New model and version: [NEW MODEL] Why I am switching: [e.g. "old model is being shut down on DATE", "lower price", "better quality"] What my application does: [DESCRIBE, e.g. "customer-support summarizer called 20,000 times a day"] How the model is called: [API, SDK, framework, agent harness] Features I depend on: [e.g. "function calling, structured JSON output, long context, vision, caching"] Constraints: [deadline, budget, compliance rules, languages] What I already have for testing: [e.g. "nothing", "50 saved prompts", "an evaluation set"] Produce: 1. Differences to check: a list of what commonly changes between model versions (output length limits, context window, pricing per input/output token, supported features, default settings, refusal behavior, formatting habits, latency). Mark each item as "verify on the provider's official page" and do not invent specifics about the models. 2. A test plan: how to build a test set of 30-100 real prompts from my traffic, which edge cases to include, and how to score answers (exact checks, rubric, human review). 3. A side-by-side comparison table template for quality, cost per task, latency and failure rate. 4. Prompt changes to try if results differ (shorter instructions, explicit output format, examples). 5. A rollout plan: shadow traffic, small percentage canary, monitoring metrics and a rollback trigger. 6. A short list of questions I should answer before starting. Rules: - Do not claim facts about specific model versions you cannot verify; tell me what to look up and where. - Keep the plan practical for a team of [NUMBER] people.
Works with: ChatGPT, Claude, Gemini, Microsoft Copilot
You are a pragmatic FinOps analyst who reduces AI API bills without hurting quality. Here is a summary of my usage over [PERIOD, e.g. the last 30 days]. For each workflow, I list the model, requests per day, average input tokens, average output tokens, whether a human waits for the answer, and the share of the prompt that is identical between requests: [PASTE USAGE TABLE] My current monthly spend is [AMOUNT]. My quality bar is: [HOW I JUDGE A GOOD ANSWER, e.g. EVAL SET PASS RATE]. I have NOT verified current provider prices. Do not quote prices; tell me which pricing page to check. Steps: 1. Rank my workflows by estimated share of spend, and say which numbers you used. 2. For the top three, evaluate these levers: prompt caching of the stable prefix, batch processing for work nobody waits on, routing simple requests to a smaller model, output length limits, lower reasoning effort and shorter context. 3. For each lever, say whether it can change answer quality (yes or no), what I should measure before and after, and the smallest safe experiment. 4. Mark anything you are assuming as "ASSUMPTION". 5. End with a one-week plan ordered from changes that do not alter outputs to changes that might.
Works with: ChatGPT, Claude, Gemini, Any LLM