Kimi K3 Burns Through $14 of Tokens in an Hour and Still Earns a Grudging Thumbs-Up
Kimi K3 delivers agentic coding capabilities that match the best closed-source Western models, but its token economics make it a liability for sustained work. Developers evaluating it need to budget for the 699-yuan tier upfront or accept that a single afternoon of heavy use will exhaust a smaller plan and lock them out for hours.
Kimi K3 approaches coding tasks by first building a miniature prototype to validate the approach, then generating production code and writing its own test cases. The output is frequently copy-paste-and-run ready, which eliminates debugging drudgery for experienced developers and lowers the barrier enough that beginners can add features without collapsing the codebase. For multi-file projects, it spins up four or five concurrent sub-agents that work in parallel, effectively acting as a temporary development team. The model is open source and its performance on frontend UI generation and backend agent orchestration already sits in the top tier globally, rivaling closed-source alternatives like Fable 5.
The cost is the catch. A 99-yuan (roughly $14) package was drained in a single hour on a moderately complex project, triggering a mandatory five-hour cooldown. Even trivial tasks trigger elaborate planning phases that burn tokens on unnecessary ceremony. Generated code runs on the first try but still requires manual review: button positions drift by two pixels, and hidden business logic gets missed. Speed varies wildly, with complex tasks slowing to a crawl that demands patience. The model is powerful but financially punishing, and it over-thinks simple requests to the point of waste.
The gap between top Chinese and Western coding agents has effectively closed on capability, shifting the battleground to pricing and token efficiency.
Kimi K3's over-planning on trivial tasks suggests its reasoning budget is not adaptive — it spends the same cognitive overhead on a one-line script as on a multi-file project, which is a design choice that directly costs the user money.
Mandatory cooldowns after exhausting a token package turn a productivity tool into a rate-limited service, which is a hard sell for professional workflows that need sustained access.
Open-sourcing a model this capable hands a turnkey AI development capability to countries and teams that cannot train their own frontier models, which has geopolitical as well as commercial implications.