Kimi K3 arrives with 2.8T params and pricey tokens

Moonshot AI has launched Kimi K3, calling it its “most capable model to date” with 2.8 trillion parameters. Announced on July 16, 2026, the model is live through Moonshot’s website and API, with an open-weight release promised by July 27, 2026.

Moonshot describes K3 as the first “open 3T-class model”, edging past DeepSeek’s 1.6T v4 Pro. In Moonshot’s own benchmark claims, K3 mostly beats Claude Opus 4.8 max and GPT-5.5 high, but trails Claude Fable 5 and GPT-5.6 Sol.

According to Artificial Analysis, K3:

Recommended reading

200 Experts Warn as AI Push Speeds Ahead

• reached an overall Elo of 1547 on its private long-horizon knowledge work evaluation
• improved by +732 points over Kimi K2.6
• ranked behind only Claude Fable 5 on that test
• cost $0.94 per task, versus $1.04 for GPT-5.6 Sol and $1.80 for Opus 4.8
• used 21% fewer output tokens on the Artificial Analysis Intelligence Index than K2.6

The model also now leads Arena.ai’s Frontend Code arena, ahead of Claude Fable 5.

Pricing is one of the more striking changes. Kimi K3 costs $3 per million input tokens and $15 per million output tokens, putting it roughly in line with Anthropic’s Claude Sonnet pricing and making it the most expensive model released by a Chinese AI lab to date, according to the source. That is a steep jump from Kimi K2.6 at $0.95/$4.

Simon Willison put K3 through his long-running “Generate an SVG of a pelican riding a bicycle” test using OpenRouter. The result consumed 95 input tokens and 16,658 output tokens, including 13,241 reasoning tokens, for a total cost of $0.25.

That led to one of the more revealing details in the post: K3 currently appears to offer only one reasoning setting, “max.” Willison notes that even a trivial prompt like “hi” counted as 86 tokens, which he says suggests there may be an 85-token hidden system prompt, though the model would not reveal it.

What the pelican test still shows

Willison argues that the pelican benchmark is no longer a reliable proxy for overall model quality. It once loosely tracked model capability, but he says that connection has now “mostly been severed.” The bigger gap is that the test says little about what matters most in current models: agentic tool calling and the ability to use tools reliably over long conversations.

Still, he says the exercise remains useful as a quick hands-on check. For K3, it surfaced a few concrete traits: high reasoning-token use, high cost for simple tasks, solid vision performance, and the ability to produce valid SVG with a reasonable grasp of geometry and spatial layout.

For Willison, that is enough to keep the pelican around—especially since comparing outputs across releases still shows progress. His verdict on this one: K3's pelican is a notable improvement over Kimi 2.5.