Kimi K3 Matches Opus 5 Quality at Lower Cost

Lin Qiao, CEO of Fireworks AI, posted results from internal benchmarks pitting Moonshot AI's Kimi K3 against Anthropic's Opus 5. Task-level quality was close across SWE, algorithmic, and terminal workloads, with Opus 5 on par or slightly ahead in some cases. Kimi K3 delivered the same work at 2x to 4.6x lower per-task cost under serverless pricing. A separate Julius harness test on a Minecraft task produced similar cost and speed advantages for Kimi K3. The comparison highlights open-model verbosity and the shift toward task-based rather than token-based evaluation.