Kimi K3 vs Claude Fable 5: benchmarks, tokens, and cost
Kimi K3 ranked first on the front-end code arena, but Moonshot says it still trails Fable and Sol overall. It used far more tokens and time on real tasks, so test it on your own work.
The main idea
On July 16, 2026, Moonshot AI released Kimi K3, a 2.8 trillion parameter model with a 1 million token context window. It reached number one on the front-end code arena, ahead of Claude Fable 5 and GPT 5.6 Sol, and ranked first in six of seven front-end categories. On Artificial Analysis it scored 57, behind Fable 5 and GPT 5.6 Sol. Kimi's own launch article says K3 still trails Fable and Sol overall.
AI Coding Daily ran 25 coding prompts and K3 scored 22 and a half out of 25, but the prompts took a full day. Chase AI gave Kimi, Fable, and Sol the same front-end task. Kimi used about 21.5 million tokens and 1 hour 33 minutes. Fable used about 3.5 million tokens and 17 minutes. Sol used about 5.5 million tokens and 25 minutes.
Price per token does not equal cost per completed task. Kimi's API price is $3 per million input tokens and $15 per million output tokens, but demand pushed its capacity close to the limit three days after launch. K3 is a mixture of experts model with 896 experts and 16 active per token. That cuts compute, but all weights must still be available. At 4-bit, the weights are about 1.4 terabytes, so you cannot run K3 on a laptop.
My verdict is that Kimi K3 is worth testing, and I would not switch blindly. Give Kimi, Fable, and Sol the same real task and record time, tokens, retries, and failures.
Key points
- Kimi K3 ranked first in six of seven front-end code arena categories, and Moonshot says it still trails Fable and Sol overall.
- On the same front-end task, Kimi used about 21.5 million tokens and 1 hour 33 minutes. Fable used about 3.5 million and 17 minutes.
- K3 has 896 experts and activates 16 for each token, which does not make a 2.8 trillion parameter model fit on a laptop.
- The 4-bit weights are about 1.4 terabytes, and Kimi recommends a super node with 64 or more accelerators.
- API pricing is $3 per million input tokens and $15 per million output tokens, with cached input at 30 cents per million.
- Full weights are scheduled for release by July 27, 2026.
Chapters
- 0:00Number one on the arena, and the caveats
- 2:25Real tests on full coding tasks
- 3:17Same task, three models: 21.5M tokens vs 3.5M
- 4:19896 experts, 16 active per token
- 6:07Open weights and what they need to run
- 10:06My verdict and how to test
Save my free seat →