Moonshot dropped Kimi K3.1 Pro into their API yesterday, and the pricing is worth paying attention to: $2 per million input tokens, $10 per million output tokens. That’s exactly what OpenAI charges for GPT-6.1 Sol and half what you’ll pay for Claude Opus 5.5.

The model isn’t open weights yet, but Moonshot says the full release happens by November 4. In the meantime, developers can hit the API and test whether the performance claims hold up in production.

Moonshot also launched K3.1 Pro Highspeed, a 2-3x faster variant at double the price for input and output. Cached input tokens cost the same across both versions, which matters if you’re doing repeated inference over the same context.

What the benchmarks show

Moonshot’s internal benchmarks put K3.1 Pro competitive with GPT and Claude on standard evals. The usual caveats apply: benchmark performance doesn’t always translate to real-world usefulness, and companies tend to publish their best numbers.

What matters more is that this pricing makes frontier-class performance accessible to smaller teams. Half the cost of Claude means you can run twice as many experiments, serve twice as many users, or just spend half as much money.

OpenAI cuts frontier models off from the internet

In other model training news, OpenAI published details about their new offline training environment for frontier models. The setup isolates model training and evaluation from the internet entirely.

Here’s what that actually means: when they’re training something like GPT-7 or running evals that might expose new capabilities, those systems can’t phone home. They can’t exfiltrate weights. They can’t call external APIs. They’re airgapped.

OpenAI says this is about containment. If a frontier model develops unexpected capabilities during training, or if red teamers find a way to jailbreak it during evaluation, the offline environment prevents that model from doing anything beyond the compute cluster it’s running on.

The practical implementation involves separate networks for different security zones, strict controls on what code can run where, and high-compute red-teaming that runs entirely offline. OpenAI can still evaluate models at scale, but those evaluations happen in containment.

This isn’t about current models like GPT-6.1. Those are already deployed, already online, already serving traffic. This is about what comes next, and what happens when you’re not sure what you’re building until it’s already trained.

The economics shift

The Kimi pricing and the OpenAI security writeup point in different directions, but they’re both responses to the same underlying reality: frontier models are getting cheaper to run and more expensive to trust.

Moonshot can undercut Claude because inference costs keep dropping. Training costs are high, but once you’ve trained the model, serving it gets cheaper every quarter as hardware improves and inference optimizations compound.

Meanwhile, OpenAI is spending engineering resources on elaborate containment infrastructure because they’re not certain what GPT-7 will be capable of when training finishes. That’s a real cost, and it’s not one that scales down easily.

For developers, this creates an interesting fork in the road. You can use increasingly affordable frontier models like Kimi K3.1 Pro for production workloads where cost matters. Or you can pay the premium for models from labs that are investing heavily in safety infrastructure, if that’s what your use case requires.

The gap between those two options is only going to widen.