Somewhere in Beijing and Hangzhou, two AI labs are playing a very expensive game of "anything you can do, I can do cheaper" — and this week the scoreboard got a lot more interesting.

The Coding Crown Changes Hands, Sort Of

Moonshot AI's Kimi K3 — a 2.8-trillion-parameter model with a 1-million-token context window — is now beating DeepSeek's V4 Pro on the Artificial Analysis Intelligence Index, scoring roughly 57 versus V4 Pro's 44, and leading rival GLM-5.2 by wide margins on shared benchmarks. Moonshot has committed to releasing K3's open weights on July 27, 2026, under a modified MIT license.

DeepSeek isn't standing still either: V4 Pro still holds the top open-weight score on SWE-bench Verified at 80.6%, and its weights have been sitting on Hugging Face under a plain MIT license since its April release — no waiting required.

Smarter, But Pricier

Here's the catch: Kimi K3's API pricing runs about $3 per million input tokens and $15 per million output tokens, versus DeepSeek V4 Pro's roughly $0.44 input and $0.87 output — nearly a 17x gap on output costs. K3 wins on raw intelligence; V4 Pro wins on "can actually afford to run this in production."

Until K3's weights land on July 27, DeepSeek remains the only trillion-scale option teams can self-host today for free, which matters enormously for anyone who'd rather own their inference stack than rent it. Once K3's weights do drop, expect a genuine price-versus-performance brawl for the open-source coding crown.

The AI arms race used to be about who had the biggest model. Now it's about who can give away the biggest model without going broke.

Source: MarkTechPost