June 12, 2026
Code Wars: Budget Edition
Kimi K2.7-Code: open-source coding model with better token efficiency
Cheap AI coder drops, and the comments instantly turn into a price-war meltdown
TLDR: Kimi K2.7-Code says it can do bigger coding jobs while using about 30% less compute than before, making it a cheaper-looking challenger in the AI coding race. The comments immediately turned into a drama-filled argument over whether expensive US models are still worth it — or just protected by data-security fears.
A new AI coding tool called Kimi K2.7-Code just showed up waving a very attention-grabbing promise: it can handle bigger software jobs while using about 30% fewer tokens than the previous version, which basically means less wasted chatter and potentially lower cost. On paper, it looks strong, landing close to — and sometimes above — some famous rivals on coding and tool-use tests. But in the court of public opinion, the benchmark charts were only the appetizer. The real main course was a community comment section turning into a live debate over price, trust, and whether premium AI is becoming a luxury handbag with a monthly subscription.
The loudest reaction was basically: "Why pay so much more for only a little better?" One commenter flat-out wondered how pricey US models stay competitive when cheaper Chinese rivals keep creeping closer in quality. That immediately opened the spicier subplot: is the real advantage not intelligence, but simply that big companies don’t want sensitive data sent abroad? In other words, is the moat just geography and compliance paperwork?
Meanwhile, practical users piled in with the most relatable energy imaginable: please, somebody tell me what actually works in real life. People asked whether setups like open-source tools plus Kimi can finally replace costly favorites like Claude Code. Others sounded almost gleefully ruthless, arguing that if a model isn’t dramatically better than cheaper options, it’s basically already obsolete. The mood was part shopper revolt, part AI horse-race, with a side of “I’m not loyal, I’m billing-conscious.”
Key Points
- •Kimi K2.7 Code is introduced as a coding-focused agentic model built on Kimi K2.6, with reported improvements on long-horizon software engineering tasks and roughly 30% lower thinking-token usage.
- •The model uses a Mixture-of-Experts architecture with 1T total parameters, 32B activated parameters, 61 layers, 384 experts, 8 selected experts per token, and a 256K context length.
- •Published benchmark results compare Kimi K2.7 Code with Kimi K2.6, GPT-5.5, and Claude Opus 4.8 across coding, ML, agentic, and tool-use evaluations.
- •Kimi K2.7 Code scores 67.4 on Kimi Code Bench v2, 63.8 on Program Bench, 42.8 on MLS-Bench-Lite, 50.4 on Kimi Claw 24/7 Bench, 81.3 on MCP Atlas, and 76.4 on MCPMark-Verified.
- •The article discloses evaluation settings, including thinking mode via Kimi Code CLI for Kimi models, Codex xhigh for GPT-5.5, Claude Code xhigh for Opus 4.8, and benchmark-specific configurations such as tool-call budgets and multi-run averaging.