Claude, Anthropic's Sonnet and Opus Families

I split my Claude work by tier. Sonnet carries the volume: drafts, routine code, the long tail of small tasks. The top tiers, Opus and now Fable, get the work where one wrong detail costs more than the tokens: architecture calls, code review, anything I would have to redo by hand if it went wrong.

What changed my setup most this year was plan economics, not a benchmark. On 10 Jul 2026 I moved from Max 20x down to Max 5x at USD 100 a month. My weekly hours on the top tiers roughly halved, so bulk and repetitive work now goes to cheaper models and Claude keeps the judgment work.

Anthropic's bias toward refusing low-confidence outputs is the right trade-off for production work. I rarely get a confidently wrong answer from it. The flip side is more "I am not sure" responses, which I have learned to read as a feature, not a bug.

The Indian-operator angle: Claude Pro is USD 17 a month on annual billing or USD 20 month to month, and Max costs USD 100 a month for 5x the Pro usage or USD 200 for 20x. At Rs 95.88 per US dollar, the rate I pulled on 28 Sep 2026, Max 5x comes to about Rs 9,588 a month. Pro is the right start if you use the model a few hours a week; Max is for all-day agent work.

What I would change is the round-trip for visuals. For images I still send the job to OpenAI's image model and bring the result back, and that hop is friction.

Use Claude when reasoning under uncertainty matters. Use it when the cost of a wrong answer is higher than the cost of a slow one. Use it when you want a model that admits its limits. Avoid it for high-volume razor-thin margin work where Haiku or a cheaper vendor model wins on cost.

Frontier models now trade places on benchmarks every few weeks. What decides my picks is how a model behaves under uncertainty and how reliably it uses tools across a long task, and that is where Claude has held its place in my stack.