The new K3 Chat & Agent mode is genuinely useful, especially for programming and development work. Combining reasoning, tools, and agent capabilities into one workflow makes complex tasks smoother.
The problem is not the combined mode itself — it is the lack of transparency and control over credit consumption.
Previously, K2.6 Thinking provided strong reasoning with tools as a normal chat mode, while heavier agent functions were clearly separate. Now every K3 conversation is billed as Agent usage, even when the user only needs a reasoning-level response.
Concrete issues:
- Users cannot see when a normal conversation escalates into a heavier agent process.
- There is no estimate or warning before an expensive task begins.
- The user cannot opt out of agent execution when it is not needed.
- Long reasoning chains, repeated tool calls, or unnecessary agent loops can consume credits without proportional value.
- Usage details show only the percentage consumed, not what caused the cost.
Requested changes, in priority order:
- A mode or switch that keeps strong reasoning and tools but disables autonomous agent execution — effectively the former K2.6 Thinking behavior. (Alternatively: separate Chat / Auto / Agent modes.)
- A per-task cost breakdown: reasoning, web search, tool calls, file processing, agent loops.
- A credit estimate or warning before high-cost tasks start.
- A configurable maximum credit limit per task.
The unified K3 experience can be excellent for serious development work, but users need to know what they are paying for and must be able to prevent runaway credit consumption. Right now the safest option for routine questions is to avoid K3 entirely, which is presumably not the outcome you want.