Guides: start from the job, not the model name

Three short, job-first guides. Each guide states what was verified from the provider page and what was not tested, so you can follow it without guessing.

Guide 1: Choose a cheap model for high-volume work

Compare Claude Haiku 5.5 ($0.10/$0.50 per million tokens under 100,000 tokens) with GPT-6 Luna ($0.10/$0.50). Rule: use the cheap model for summaries, classifications and subagent work; escalate to a flagship only when the cheap model’s output quality fails a specific, checkable test on your own data. Do not escalate on a benchmark headline alone.

Guide 2: Budget a flagship without a December surprise

Use Gemini 4 Argon as the worked example: budget at $4/$20 per million tokens, the post-30-November price, and treat $2/$10 as a trial window. If your workload only fits at the introductory price, your plan is not yet viable — change the workload, the model, or the price assumption, in that order.

Guide 3: Use an open-weight model without a licence surprise

Use FLUX 3 Image Klein and Muse Spark Open as the worked examples: check the licence on the official website for your revenue band and your intended use (client work, redistribution, fine-tuning) before delivery. “Downloadable” is not the same as “unrestricted commercial use”.