Moonshot AI's new Kimi K3 model narrows the gap with America's top labs, but the real story is the price tag. China just stopped competing on discount.
Moonshot released Kimi K3 on July 16, a 2.8 trillion parameter open model built with a new Kimi Delta Attention mechanism and a million-token context window. Bloomberg's framing was blunt: the model closes the gap with US rivals. The launch reportedly rattled AI and chip stocks, with investors drawing comparisons to last year's DeepSeek moment. But the number that matters most sits underneath the benchmark charts. Moonshot priced K3 at $3 per million input tokens and $15 per million output tokens.
That's roughly five times what the Kimi K2 family cost. It's also, according to the-decoder, the highest list price any Chinese lab has ever attached to a model. For a year and a half, the pitch out of Beijing has been consistent: match Silicon Valley's capability, undercut it on price. DeepSeek built its entire identity on that trade. Kimi K3 abandons it. At $3/$15, K3 lands in the same pricing tier as Anthropic's Claude Sonnet 5, not in DeepSeek's bargain bin, where V4 Pro still runs about 8 cents on the dollar by comparison on cost-per-task metrics.
How the Price Stacks Up #
Here's the thing. K3 still costs less than the frontier models it's chasing. Claude Opus 4.8 runs $5 per million input tokens and $25 per million output tokens, so K3 undercuts it by roughly 40%. Against OpenAI's GPT-5.6 Sol, priced at $5 and $30, the gap is even wider. Artificial Analysis puts K3's average cost per task on its Intelligence Index at $0.94, close to Sol's $1.04 and about half of Opus 4.8's $1.80. The math still favors K3.
On raw capability, Moonshot's own benchmark table shows K3 leading in seven categories, including BrowseComp, FrontierSWE, GDPval-AA and Toolathlon. Opus 4.8 still wins on GPQA, Humanity's Last Exam and OfficeQA Pro. Sol edges K3 on Terminal-Bench, 88.8% to 88.3%. That's thin enough to argue about, but real enough to matter to anyone routing coding workloads by benchmark score. Moonshot itself concedes K3 still trails Claude Fable 5 and GPT-5.6 Sol on overall performance. It's close. It's not first.
Independent testing backs some of the buzz. In blind evaluations run by Arena, developers reportedly preferred Kimi K3 over every leading US model, Fable 5 and GPT-5.6 Sol included, specifically for front-end coding tasks. In Arena's broader text ranking, K3 finished ahead of Opus 4.8 while costing 40% less to run. Axios called it plainly: China just erased America's AI lead. That's a strong claim for one launch, and it deserves some skepticism. But the benchmark spread is narrow enough now that the claim holds up on the numbers.
What Comes Next #
What makes this launch different from the last four Kimi headlines isn't the score, it's the pricing signal. Moonshot could have shipped K3 at K2-era rates and let the capability jump alone do the talking. It didn't. Charging flagship-adjacent prices is a bet that its 2.8 trillion parameters are worth paying for. Cheap alone was the old pitch. That's a different kind of confidence than the one Chinese labs have projected since DeepSeek's debut.
Moonshot plans to release K3's open weights on July 27, letting companies and governments run and customize the model on their own infrastructure rather than renting API access. That detail matters more for Washington than the benchmark table does. A model that rivals Opus 4.8 on cost and text ranking, and that anyone can download and self-host in under two weeks, is a harder thing to regulate away than a hosted API. The price war may be over. The infrastructure question is just getting started.
Also read: Kimi K3 Shows the Cheap Chinese AI Era Is Coming to an End • Upper90 Lends $400 Million Against AI Inference Chips Instead of Nvidia GPUs • Xi Jinping tells the world AI should be a symphony, not a solo performance