Every AI codebase you ship has the same problem lurking inside it. Somewhere in your settings file or your prompt loader, a specific model name is hardcoded. gpt-4o here. claude-3.5-sonnet there. deepseek-v3 in the one service you built last month.
Six months from now, at least two of those names will be deprecated. The other one will be dominated by a newer, better, cheaper version of itself that you have not upgraded to because nobody remembered which service was using which model.
That is the model naming trap. It is entirely avoidable.
Names describe supply. Tiers describe demand.
A model name tells you what the vendor is selling right now. It tells you nothing about the shape of your task or how much quality it deserves.
A tier profile flips this. Instead of saying "route this call to Claude Sonnet 5", you say "route this call to whatever the current balanced default is". The tier is a promise about capability and cost. The name is a moving target the platform maintains for you.
When Sonnet 6 ships next quarter, your Flow tier upgrades. Your code did not change. Your bill went down or your quality went up, depending on what the platform decided to route you to.
Three tiers cover almost everything
Most workloads sit in one of three profiles.
Core is fast and cheap. Haiku class. Use it for classification, extraction, high volume agent steps, and anything where speed and cost dominate quality.
Flow is balanced. Sonnet class. This is the sensible default for real product work, drafting, multi tool orchestration, and anything that runs long enough for cost to matter.
Max is deep reasoning. Kimi K3 with a Sonnet fallback when the request has images. Use it for research, code review, strategy, and consequential decisions where being wrong is expensive.
There is a fourth tier, Auto, that is not really a tier at all. It is a small router that reads your prompt shape (length, tools, images, reasoning cues) and picks Core, Flow, or Max on your behalf. For most calls, this is what you want.
The economics nobody talks about
The naive strategy is to always route to the best model you can afford. On a call by call basis this looks smart. Across ten thousand calls it is expensive theatre.
At current prices, running everything through a top tier model costs roughly ten times what routing intelligently costs. The quality delta on eighty percent of your traffic is single digit percentage points. On the twenty percent where quality matters, you want the top tier and Auto sends you there. On the rest, you are burning credits for nothing.
We built the T1 tier system because we got tired of watching teams pay Opus prices for Haiku work.
The one line change
If you are already using the OpenAI SDK or Anthropic SDK, T1 works as a drop in.
Point at engine.taskclan.com/api/openai/v1 or engine.taskclan.com/api/anthropic/v1 as your base URL. Use sk_t1_ as your key. Set the model to t1-auto and forget about it.
Or use the Taskclan SDK directly, and call taskclan.t1.chat({ tier: 'auto', messages }). Same result, cleaner surface.
Full docs on tiers at taskclan.com/docs/models. Playground at platform.taskclan.com/playground if you want to try it without a single line of code.