Before Kimi K2.6 appeared, teams ran workloads on Claude Opus 4.7 at $25 per million output tokens. On 20.04. the open model Kimi K2.6 was released and immediately changed cost dynamics.

What Kimi K2.6 is

Kimi K2.6 is an open Mixture-of-Experts model with a theoretical size of 1 trillion parameters and 32 billion active parameters per token. Its inference behaves like a compact 32B model while delivering performance comparable to a 1T architecture.

Costs and benchmarks

The published pricing is $0.60 per million input tokens and $2.50 per million output tokens, with a 256K token context window. On SWE-Bench Pro Kimi scored 58.6%, reportedly outperforming GPT-5.4 and Opus 4.6.

For an agent cycle with 100 million input and 10 million output tokens monthly, Opus 4.7 costs $2,550 while Kimi K2.6 costs $85. That gap implies an annual saving of $28,560 on identical workloads.

Real-world tests

Independent tasks demonstrate Kimi’s strengths in long autonomous runs. In one Mac assignment to build a Zig engine from scratch, Kimi completed 14 iterations and 4,000+ tool calls in 12 hours, producing a result about 20% faster than LM Studio.

In another case, Exchange-Core, an open-source financial engine with eight years in production, Kimi acted as senior architect for 13 hours, producing 4,000 lines of changes and nearly tripling throughput on critical paths.

Recommended orchestration and pitfalls

Most teams currently rely on a single model. The advised stack is to use Kimi K2.6 as the main brain for 80% of tasks, keep Opus 4.7 for the top 10% of complex escalations, and use Haiku 4.5 for headlines and session compression.

Kimi’s thinking mode generates about 3.6x more output tokens than Opus on similar tasks; with tokens priced roughly ten times lower, the effective saving becomes approximately 2.7x, not tenfold. Cheap tokens also reduce monitoring discipline, which can lead to unnoticed extra consumption and occasional 80K token overruns.

«Scope: [specific files]. Do not touch anything outside the scope. List related issues at the end, do not fix inline.»

Conclusions

Kimi K2.6 does not fully replace Opus for every use case, but it enables shifting most workloads to a far cheaper open model. The model is open-source and is expected to spawn forks and specialized versions in months, expanding options for deployment and optimization.