Kimi K2.7-Code

“The price and the weights are real; the capability claims are homemade. Buy it as a value play, never as a frontier substitute.”

Strategically this is a hedged bet: open weights eliminate lock-in, the price creates immediate savings on suitable workloads, and Moonshot has shipped a coherent K2 line at a steady cadence. But the disputed benchmark situation is a governance signal, not just a research footnote — a vendor whose numbers practitioners publicly challenge within days is a vendor whose future claims you must independently verify, forever. That verification tax belongs in the TCO. Position it for cost-tiered coding traffic behind your own eval gate, with a frontier model above it for the work that matters.

“Moonshot is racing GLM for the open coding-agent market on price and modality — and litigating capability in benchmarks nobody else can run.”

The positioning is sharp: while GLM-5.2 owns the open-weights capability crown, K2.7-Code differentiates on price, video-input modality, and thinking persistence — three axes that matter to coding-agent builders specifically. Launching vendor-designed benchmarks (Kimi Code Bench, MCP-Atlas variants) is a deliberate attempt to define the scoreboard, and the practitioner backlash shows the market isn't buying it. The durable wedge is multimodal coding: nobody else open ships video-to-code. If Moonshot lets independents validate the next release, the trust gap closes; K3 is already slated for Q3 2026.

“A fifth of frontier coding costs, and caching takes it lower — the eval tax to trust it is the only real line item.”

The raw economics are excellent: $0.95/$4.00 list, $0.19 cache hits, OpenRouter blended around $0.74/$3.50, and Moonshot's caching design fits agent workloads where context repeats — effective costs 60-80% under list are realistic. Always-on thinking cuts the other way: you pay for reasoning tokens on every call, and the claimed 30% reduction vs K2.6 is a vendor number like the rest. Self-hosting with INT4 is viable at scale. Budget one real cost the rate card hides: engineering time to build the task-level evals that substitute for the missing trustworthy benchmarks.

“Long agent runs hold together impressively — the persistent thinking is a real feature. Just don't expect the model the launch charts promised.”

Hands-on, the persistent-thinking design is the standout: multi-turn agent sessions keep their reasoning thread instead of re-deriving state, which shows up as fewer dropped constraints on long tasks. Tool invocation is solid; tool_choice limited to auto/none and the fixed sampling are mild annoyances. Screenshot-to-code works; video-to-repro (API-only) is genuinely novel. Against GLM-5.2 or frontier coders on hard multi-file work, it loses more often than the vendor charts imply — consistent with the practitioner pushback. The 32K output ceiling forces chunked generation on big diffs. Set expectations one tier down from the marketing and it's a satisfying tool.

“For the price it's a lovely daily coder — visible thinking, reads my screenshots, occasionally watches my screen recordings. Frontier it is not.”

As an everyday coding companion via OpenRouter or the first-party API, K2.7-Code delivers more than its price suggests: always-visible reasoning you can audit, competent multi-language code, and the party trick of accepting a screen recording of a bug and producing a plausible reproduction. The highspeed variant at ~180 t/s keeps interactive flow pleasant. Rough edges: the mandatory thinking makes trivial questions slower than they should be, output caps at 32K, and on genuinely hard problems it noticeably trails Claude/GPT/GLM-class models. English and Chinese are both strong.

“Vendor-designed benchmarks, vendor-run scores, public practitioner pushback within days — this launch is a case study in why we verify.”

The core problem: Moonshot's headline numbers live on suites Moonshot created (Kimi Code Bench v2, where its own chart shows GPT-5.5 ahead at 69.0 vs 62.0 — at least that's candid), with no independent replication path, and VentureBeat documented practitioners publicly disputing that the claimed performance holds on real tasks. The only third-party SWE-bench figure (~60.4, informal harness) sits far below the frontier and below what the marketing implies. Add an undisclosed knowledge cutoff, no safety documentation, forced thinking you cannot audit off, and a "modified" MIT license, and the pattern is consistent: maximum control of the narrative, minimum external verifiability. The model underneath appears genuinely useful at its price — which makes the benchmark theater unnecessary and the trust damage self-inflicted.