Kimi K3 entered the market with a bold claim: number one on the Arena coding leaderboard, ahead of Claude Fable 5 and GPT-5.6 Sol. Then the pricing landed-output at 100 CNY per million tokens, up from K2.6's 27 CNY.
For product teams, this is not a pricing complaint. It is a routing decision. Most AI product teams picked one model and built around it. The token rate was a line item. When a new model arrived, you either switched or you did not. The decision was binary.
K3 breaks that assumption. K2.6 is still available and cheaper. K3 is more capable but 3.5x more expensive. Both come from the same provider, with the same API surface. The question is no longer "which model?" but "which model for which task?"
Instead of choosing one model, build a routing layer that selects based on task characteristics:
| Task type | Difficulty signal | Routed model | Rationale |
|---|---|---|---|
| Simple code completion | Short context, single function | K2.6 | Low cost, sufficient quality |
| Multi-file refactoring | Large diff, cross-module | K3 | Higher first-pass accuracy justifies cost |
| Bug diagnosis | Ambiguous, needs reasoning | K3 | Arena-leading reasoning reduces retries |
| Boilerplate generation | Template, repetitive | K2.6 | Marginal quality difference, cost dominates |
| Architecture review | Complex, high-stakes | K3 | Error cost exceeds token cost |
Per-token cost is the wrong unit for product decisions. What matters is the total cost of producing a task outcome your team accepts and ships.
If K3 gets a refactoring task right on the first pass and K2.6 needs three retries, K3 may be cheaper despite costing 3.5x per token. If K2.6 handles boilerplate fine, routing it to K3 wastes money. K3 was so popular that Kimi suspended new subscriptions within 48 hours of launch. The cluster could not handle the load. That tells you two things:
For product teams, this means your own infrastructure decisions matter. If you route everything to the best model, you may hit rate limits or face suspended access. A routing layer that falls back to a cheaper model for simple tasks protects you from outages. This is a decision framework, not a benchmark. I have not measured K3 against K2.6 on specific tasks. The routing table above is a hypothesis based on the Arena ranking and pricing data, not verified results. You need to run your own task-level comparison before committing to a routing strategy.
The framework also assumes both models are available. As of 2026-07-21, K3 subscriptions are d. Your routing layer needs a fallback plan.
Disclosure: I'm a MonkeyCode user sharing my own experience, not affiliated with the project. MonkeyCode is an open-source AI coding platform: https://github.com/chaitin/MonkeyCode