Peter Smith
⬤ Kimi.ai just dropped Kimi K2.5, their latest open-source visual and agentic AI model built for heavy-duty reasoning, multimodal work, and running multiple agents at once. The launch puts serious emphasis on two things: solid benchmark numbers and pricing that undercuts most of the competition, making Kimi K2.5 a budget-friendly option against the big proprietary players.
⬤ The benchmark results look pretty competitive across the board. Kimi K2.5 hit 50.2 percent on Humanity's Last Exam Full, 74.9 percent on BrowseComp, and managed 76.8 percent on SWE-Bench Verified. It also pulled 78.5 percent on MMMU Pro, 86.6 percent on VideoMMMU, and 88.8 percent on OmniDocBench 1.5. Depending on what you're testing, these scores put it right alongside or even ahead of models like GPT-5.2 xHigh, Claude Opus 4.5, and Gemini 3 Pro.
⬤ There's also a new Agent Swarm feature designed to run multiple agents in parallel. The system uses an orchestrator that coordinates specialized sub-agents—think AI researchers, fact checkers, web developers, and niche domain experts all working simultaneously. It can handle up to 100 sub-agents and process as many as 1,500 tool calls in one workflow, which makes tackling complex multi-step projects way more efficient than traditional single-agent setups.
⬤ The pricing structure is where things get interesting. Kimi K2.5 costs just $0.60 per million input tokens and $3.00 per million output tokens—substantially cheaper than most models performing at similar levels. You can access it now on kimi.com in both chat and agent modes, with Agent Swarm currently in beta for premium users. It's a clear signal that the company's betting big on scalable, open-source agentic AI with a strong price-to-performance ratio.
Peter Smith
