Article: Moonshot AI’s new Kimi K3 topped GPT-5.6 on the AA-Briefcase benchmark, scoring 1,527 against the latter’s 1,495. The win comes with a pricing model that makes the 2.8-trillion-parameter open-weight model a surprisingly cheap option for long-form coding tasks.
Why the benchmark matters
AA-Briefcase measures performance on sustained, real-world workloads rather than isolated trivia questions. A higher score means a model can keep context, plan ahead and stay on task over thousands of tokens – exactly the conditions developers face when building assistants, code generators or research aides. Kimi K3’s edge over GPT-5.6 shows an open model can now compete where closed rivals have traditionally ruled.
The technical picture
- Size – 2.8 trillion parameters, the largest open-weight model released to date.
- Architecture – Stable LatentMoE (Mixture-of-Experts) with 896 experts, of which 16 are active at any moment.
- Context window – Over 1 million tokens, enough to hold entire books or long codebases.
- Attention – Kimi Delta Attention, a custom tweak claimed to improve scaling efficiency.
These choices give the model the capacity to handle “long-horizon” tasks, but they also drive up compute requirements, which shows up in speed and cost figures.
Pricing that catches the eye
Moonshot breaks token pricing into three buckets:
| Usage type | Cost per 1 M tokens |
|---|---|
| Input – cache miss | $3.00 |
| Input – cache hit | $0.30 |
| Output | $15.00 |
The company says coding workloads see a 90 % cache-hit rate. If true, it is a very cheap option for coding agents.
The trade-offs you’ll feel
- Speed – The model processes about 62 tokens per second, below the field median. A long prompt can linger for minutes, which matters for interactive use.
- Reasoning cost – Kimi K3 runs with “max reasoning effort” by default and offers no knob to dial it back. Simple queries therefore burn the same token budget as complex ones, inflating costs for routine tasks.
- Hidden token usage – Some users report a sizable system prompt that the model injects automatically, adding tokens that are billed but not visible in the user’s request.
These factors mean the headline-low price can be offset by slower turnaround and higher token consumption in practice.
Availability and the road ahead
The API is live today, letting developers experiment immediately. The model weights themselves will be released on July 27, after which self-hosting becomes possible. Until then, users must rely on Moonshot’s hosted service and accept the associated latency and pricing structure.
Counter-point from the closed-model camp
The emerging takeaway
The main takeaway is the shrinking gap between closed models and open models. Kimi K3 proves that open-weight models can handle complex, long-horizon jobs.